Abstract
An increasing number of studies have used multi-armed bandit tasks to investigate individual differences in exploration behavior. However, the psychometric properties of exploration measures remain unexplored. We examine the test-retest reliability, convergent, divergent, and external validity of model-based estimates of exploration strategies using three canonical paradigms. Our results revealed poor to moderate reliability, with minimal correlations for the same strategy across tasks. We then provide actionable recommendations for how to improve reliability and convergence across tasks: Simplifying common computational models enabled us to identify two convergently valid latent factors representing value-guided and directed exploration. Still, these factors showed neither a significant correlation with self-reported exploration tendencies nor with mood fluctuations, symptoms of anxiety, and depression. The exploration factors were, however, highly correlated with working memory capacity, questioning whether they provide additional information beyond performance-related constructs. To improve future research, we suggest simplifying common computational models and using multiple tasks to more accurately measure exploration strategies and mitigate spurious correlations arising from task-specific factors.