Thank a lot! That is very interesting indeed. We also found that across all tasks the model-free switch probability (probability of choosing a different option than on the previous trial) is more reliable and has greater convergent validity. Very happy to discuss more about this.

This work was pregistered with the OSF and all code and data is publicly available.
Preregistration: osf.io/cavj3
data and code: osf.io/ra7su/

In sum, we show that simplified modelling and creating latent factors improves robustness. Still, extracted strategies are more linked to working memory than real-world exploration, raising questions about the validity of few-armed bandit tasks for studying exploration.

Our changes improved the convergence of exploration strategies across tasks. The latent factors were however still not related to any self-reported measures of exploration or to any psychiatric constructs. They did however show a strong correlation with working memory capacity.

We subsequently made two main changes to the analyses: 1) We simplified and unified the computational modelling for the Horizon task and the 2-armed bandit. 2) We constructed theoretically informed latent constructs for the exploration strategies across all tasks.

When using the standard modelling approaches, the test-retest reliability of all model parameters and most task measures was rather poor. We also found low correlations between model parameters from different tasks, despite these parameters measuring the same strategies.

We tested these requirements by collecting data on three bandit tasks, five questionnaires and three working memory tasks. We retested participants on all tasks after 6 weeks, to test for temporal stability.

This type of approach is very valuable. However, for treating individual differences in model parameters as cognitive traits, we need the model parameters to be:
🕐stable over time
🔗converging across tasks
📋related to real-world exploration

A growing number of studies uses few-armed bandit tasks to test how people explore. Recently, an increasing body of research (including our own) turned to individual differences in this behaviour and related model parameters to questionnaire measures of psychiatric traits.