There are many studies that focus on proximate mechanisms of multisensory integration and some that focus on ultimate issues such as how such signals constrain speciation, but there is no overarching framework to predict when stimuli from different modalities should be integrated. We developed a model to predict when information from two different sensory modalities should be integrated. In some situations, using a single stimulus is advantageous over integration. We empirically evaluated key predictions from the model in an antipredator context using yellow-bellied marmots (Marmota flaviventris). Individuals were presented first with coyote urine followed by coyote vocalizations. In agreement with the predictions from the model, we found that marmots only integrated stimuli when foraging under greater predation risk or when the vocalizations were embedded in noise. We also tested whether integration varied under different costs of mistakes. It did not. This finding was somewhat expected given that the cost of mistakes was the least influential parameter determining integration in our model. Overall, our model makes clear predictions about the extent to which various factors influence integration with the aim of providing testable hypotheses to guide future empirical work.