Abstract
Recent research suggests that responses during experimental fear conditioning follow distinct latent trajectory classes. Computational reinforcement learning models can be used to infer latent learning processes during the task (e.g., learning rates and prediction-error sensitivity) that may help explain these classes. The present study re-analysed previously published and unpublished datasets to examine patterns of fear learning using skin conductance responses (SCRs) in a large sample of healthy participants. Data were aggregated from 12 fear conditioning experiments (N = 1,083; Mage = 20.72 years), encompassing acquisition, extinction, and return-of-fear phases. Using k-means clustering, we identified solutions comprising either two or three distinct response profiles. Across clusters, approximately half of participants showed minimal differential SCRs, while the remaining participants exhibited higher conditioned responding during acquisition, with subsets differing in whether fear returned following extinction. Cluster membership was significantly predicted by gender, current anxiety levels, post-acquisition conditioned stimulus pleasantness ratings, and Bayes factors quantifying evidence for model-based learning. Comparison of reinforcement learning models indicated that a latent cause model provided a superior account of SCR data relative to Rescorla–Wagner, Pearce–Hall, and hybrid models, with latent cause parameters differing by cluster and gender. The clustering data provide preliminary evidence that SCR outcomes in fear conditioning studies may be disproportionately influenced by a minority of participants with specific pre-existing characteristics, rather than reflecting uniform learning processes. Moreover, the results provide support for latent cause theories of fear learning and highlight the value of combining clustering and computational approaches to better understand individual differences in fear acquisition, extinction, and relapse.