Homophily, the tendency of people to interact with people who are “similar” to themselves, is a well-documented phenomenon, and can have an important effect on economic outcomes. In this paper, we provide a belief-based explanation for homophily. Our theory is based on the notion that people are better at predicting the actions of people that are similar to them, consistent with evidence from neuroscience and social psychology. Each player receives a signal what action is appropriate in a particular situation, with signals being correlated within but not between groups. Players are introspective in that they understand that others have received a signal, understand that other players understand that, and so on. We assume that group membership is unobservable, so that  players need to signal their identity using markers. We show that there is partial or complete segregation into groups if and only if there is some asymmetry in preferences over markers across groups. Notably, our introspective process predicts a unique outcome, in contrast with the usual equilibrium analysis. Moreover, while many equilibria are socially wasteful, with lock-in on inefficient outcomes, the outcome we predict is always efficient. We contrast our predictions with those from models based on social preferences and we explore some experimental implications. (Joint with Alvaro Sandroni)