Roll-call cohesion scores are the most widely used measures of voting blocs in legislative studies, appearing in literally hundreds of studies since their introduction in 1924. Despite a staple of legislative studies, we know virtually nothing about the statistical properties of these scores. In this article, it is shown how such scores suffer a serious bias problem: scores are artificially inflated for small parties, especially those that are less unified. The problem is demonstrated and an intuitive solution proposed. It is illustrated with data from the United States and from Brazil.