This paper uses data consisting of agents’ strategically reported preferences as well as their underlying true preferences to study strategic behavior in the course allocation mechanism used at Harvard Business School. We show that the mechanism is manipulable in theory, manipulated by students in practice, and that these manipulations cause meaningful welfare losses. However, we also find that ex-ante welfare is higher than under the random serial dictatorship (RSD), which is the only known mechanism that is anonymous, strategyproof and ex-post efficient. This discrepancy between ex-ante and ex-post performance of RSD is specific to the multi-unit assignment problem and can be traced to the callous behavior induced by RSD. We draw lessons for the design of multi-unit assignment mechanisms and for market design more broadly. (With Eric Budish, Chicago Booth School of Business)