Roles are a pervasive part of our social landscape, but little is known about the mental models people use to reason about agents who occupy roles. In this paper, we test three computational models for role-based reasoning against participant performance in a social inference task. We find evidence that people exhibit mixed approaches which broadly track the computational efficiency of simpler models, but still retaining the structure of Bayesian inference models. These findings shed light on the mechanics of this important social cognitive system and pave the way for future work in this area.