We develop a structural model to address the question whether, and to what extent, expert panelists engage in herd behavior when voting on important policy questions. Our data comes from FDA advisory committees voting on questions concerning the approval of new drug applications. We utilize a change in the voting procedure from sequential to simultaneous voting to identify herding. Estimates suggest that around half of the panelists are willing to vote against their private assessment if votes from previous experts indicate otherwise and, on average, 9 percent of the sequential votes are actual herd-votes. Temporary committee members are more prone to herding than regular (standing) members. We find that simultaneous voting improves information aggregation given our estimates.