A well-tested hypothesis regarding how the brain might integrate multi-sensory cues to generate a coherent percept is that the cues are combined linearly, weighted by their reliability. However, a potential form of non-linearity arises when the cues suggest very different estimates of a stimulus variable. For example, when visual and auditory cues in a scene are discordant, it can result in illusory percepts such as the ventriloquism effect. Linear approaches fail under such circumstances. Following the intuition that multiple causal factors can give rise to a percept associated with the cues, recent modeling studies have proposed a class of mixture models for statistically optimal sensory combination by considering two hypotheses — either that all the cues have a common cause or that the causes are independent. Employing one such causal generative model, we test the explanatory power of Bayesian model selection and model averaging approaches in characterizing nonlinear cue combination behaviors in an auditory localization task, where human subjects are asked to report on whether auditory and visual stimuli delivered at a variety of spatial disparities are perceptually unified.
