The 2000 presidential election focused attention on the problem of unrecorded votes, in which a person casts a ballot but fails to record a valid vote for a particular contest. Although much recent research has evaluated voting technologies and their effects on unrecorded votes, there has been little research on the effects of ballot design. We argue that the same theories used to design and evaluate self-administered surveys can be used to analyze ballot features. We collect and code paper-based ballots used in the 2002 general election from 250 counties in 5 states. We code the ballots in terms of several graphic design elements, including the content and location of ballot instructions and the layout of candidate names and office titles. Our analysis suggests that several ballot features are associated with unrecorded votes (both overvotes and undervotes) in the gubernatorial contests. We also find that ballot design features exacerbate the racial disparity in unrecorded votes. Ballot design can be an important factor in determining whether voters are able to cast a ballot accurately, which can influence the legitimacy of elections.