Bayesian analysis of simulation-based models

Recent advancements in Bayesian modeling have allowed for likelihood-free posterior estimation. Such estimation techniques are crucial to the understanding of simulation-based models, whose likelihood functions may be difficult or even impossible to derive. One particular class of simulation-based models that have not yet benefited from the progression of Bayesian methods is the class of neurologically-plausible models of choice response time, in particular the Leaky, Competing Accumulator (LCA) model and the Feed-Forward Inhibition (FFI) model. These models are unique because their architecture was designed to embody actual neuronal properties such as inhibition, leakage, and competition. Currently, these models have not been formally compared by way of principled statistics such as the Bayes factor. Here, we use a recently developed algorithm-the probability density approximation method-to fit these models to empirical data consisting of a classic speed accuracy trade-off manipulation. Using this approach, we find some discrepancies between an assortment of model fit statistics. For some participants, one model appears to be superior when one fit statistic is used, while another appears superior when a different statistic is used. However, for 13 of the 20 participants, one model wins by all of the fit metrics considered. The FFI wins in 5 of these cases, while the LCA wins, often by a wide margin, for the others.