We develop a Roy model of social interactions in which individuals sort into peer groups based on comparative advantage. Two key results emerge: First, when comparative advantage is the guiding principle of peer group organization, the effect of moving a student into an environment with higher-achieving peers depends on where in the ability distribution she falls and the effective wages that clear the social market. In this sense our model may rationalize the widely varying estimates of peer effects found in the literature without casting group behavior as an externality in agents' objective functions. Second, since a student's comparative advantage is typically unobserved, the theory implies that important determinants of individual choice operate through the error term and may, even under random assignment, be correlated with the regressor of interest. As a result, linear in means estimates of peer effects are not identified. We show that the model's testable prediction in the presence of this confounding issue--an individual's ordinal rank predicts her behavior, ceteris paribus--is borne out in two data sets.