We quantify the effect of pre-vote deliberation on the decisions of US appellate courts. We estimate a model of strategic voting with incomplete information in which judges communicate before casting their votes, and then compare the probability of mistakes in the court with deliberation with a counterfactual of no pre-vote communication. The model has multiple equilibria, and judges' preferences and information parameters are only partially identifed. We find that deliberation can be useful when judges tend to disagree ex ante and their private information is relatively imprecise; otherwise, it tends to reduce the effectiveness of the court.