Legislative politics scholars rely heavily on roll call vote (RCV) data. However, it has been claimed that strategic motives behind RCV requests lead to overestimating party group cohesion and, thus, biased findings on legislative behaviour. To explore this claim, we distinguish between two types of bias, a ‘behavioural bias’ and a ‘selection bias’. A recent rule change in the European Parliament, making RCVs mandatory on all final legislative votes, presents the unique opportunity to evaluate the latter. We compare party group cohesion in requested and mandatory RCVs by examining final legislative votes before and after the rule adoption using amendment RCVs (which still need to be requested) as a benchmark. The analysis shows that group cohesion is higher whenever RCVs are not just requested on some but mandatory on all votes. Hence, there is indeed a ‘selection bias’ in RCV data. Yet, somewhat contrary to former claims, relying on requested RCVs leads to underestimation of the cohesion party groups would have had were all votes automatically roll called. We argue that this is mainly because requests occur on more contentious votes.