This paper develops a consensus voting model for estimating preferences of federal circuit court judges. Unlike standard ideal point models, which assume that judges vote sincerely for their preferred outcomes, the consensus model accounts for the norm of consensus in the courts of appeals by including a cost of dissent in the judicial utility function. A test of the consensus voting model on a data set of asylum appeals demonstrates that it provides a substantially better fit than a comparable sincere voting model and also generates more accurate predictions of voting probabilities. The model generates credible estimates of the impact of panel composition on case outcomes, which is surprisingly large in the asylum cases. Even though 95 percent of these decisions were unanimous, roughly half of the cases could have been decided differently if assigned to different panels.