A Bias in European Parliament Roll-Call Votes?

The potential selection bias in EP RCVs has long been neglected. Carrubba et al. (2006) were the first to address it by analyzing the whole sample of votes taken in the EP between July 1999 and 2000. They found differences between party groups concerning their RCV requests, with the EPP and the Greens being the groups asking for roll calls most often. Furthermore, Carrubba et al. discovered that votes in some issue areas were subject to roll calls much more frequently than issues in other areas. The bulk of RCVs in the period 1999-2000 were linked to subjects concerning Justice and Home Affairs, Constitutional Affairs, and Economic and Monetary Affairs. While these subjects together accounted for less than one-third of all votes, they accounted for nearly two-thirds of all RCVs. These issue areas were thus overrepresented in RCVs. Moreover, Carrubba et al. showed that legislative votes were massively underrepresented: While around 50 % of all votes on (non-legislative) resolutions became RCVs, only 7 % of all votes on legislative proposals were held by roll call. Based on these findings, Carrubba et al. argued that RCVs are subject to a selection bias. Furthermore, by assuming that EPGs use roll calls strategically to enforce cohesion among their MEPs, they inferred that RCV analysis overestimates cohesion within EPGs.

This last point, the question of whether EPG cohesion is higher in RCVs than it might be in votes that are not recorded, has raised further interest. Janina Thiem (2006) investigated strategic reasons behind RCV requests. She theorized that party group leaders are motivated by one of the following: (1) disciplining their own MEPs, (2) publicly expressing a certain policy positions, or (3) revealing the voting behavior of another EPG. Moreover, she studied the ability of party groups to sanction their members for divergent voting behavior. Finding that EPGs have nearly no means to sanction their members, she assumed that the high levels of EPG cohesion found in RCVs are due to the fact that party leaders only call RCVs on votes where they expect their MEPs to display a cohesive position either to express a certain policy position or to reveal the divergent position of another EPG (Thiem 2006, 2007)—an assumption which was supported by Stefan Tierse (2016). Carrubba et al. (2008) developed a game theoretical model of RCV selection that offers testable predictions about the selection process and about a potential selection bias. Their model implies that the observed cohesion in RCVs depends on the size and heterogeneity of the parties and on the status quo. They also provided some initial empirical tests of their model (Carrubba et al. 2009). However, it is not possible to deduce a global conclusion on a bias in RCVs from their model since their hypotheses relate only to very specific circumstances. Besides, even if we assume that EPG leaders request RCVs to discipline their rank and file, it cannot be concluded that EPG cohesion in RCVs is higher than in non-RCVs. While on the one hand, the fact that voting behavior is recorded might force group members to stick to the common line, on the other hand, RCVs mostly take place for decisions where group leaders feel that more discipline is needed because the issue at hand is particularly salient or contested or because they perceive group cohesion to be too low (Yordanova and Muhlbock 2015). Hence, there might be more cohesion in RCVs, less cohesion in RCVs, or these two opposing effects might simply cancel each other out (Hug 2010). In any case, considerable caution in the analysis of RCVs is necessary. Hoyland (2010) demonstrates that results based on RCVs vary between legislative and non-legislative procedures. Therefore, it is particularly important not to mix legislative and non-legislative votes together. Otherwise, overrepresentation of non-legislative votes in roll calls might cause a bias.

To sum up, research on RCV requests has shown that one needs to consider the following potential problems of data based on EP roll calls: ^[1]

We will now examine whether the problems with RCV data outlined above also affect the present study. As described above, we only use a subset of all recorded votes in the period from 1999 to 2009, namely only votes on Codecision legislation, votes on whole proposals, and votes from the last reading before the proposal was finally adopted by the Council and the EP. The sample thus consists of 190 RCVs. This amounts to 23.63 % of the total number of votes (on whole proposals in the last reading before final adoption) during that period (see Table 2.1). Compared to the finding of Carrubba et al. (2006) that only 7 % of all votes on legislative proposals were held by roll call, this is a fairly good ratio. However, one has to note that most of the RCVs used in the present study occurred between July 2008 and June 2009, the last year of the sixth EP.

Why is this the case? In 2008, the Independence/Democracy group (ID) started calling RCVs on nearly every vote. While the group claimed that this was done to enhance the transparency of the EP, the other MEPs were quite annoyed, because holding a roll call on every vote slowed down the procedure and the already tiring voting sessions became even longer (Interview 7). However, after negotiations between the groups, a compromise was reached. The EP changed its rules of procedure so that, since July 2009, every final vote becomes a RCV (cf. Rule 166 of the EP’s rules of procedure⁷ and Interview 10). This has three implications: First, the dataset is fortunately larger than could have been expected according to Carrubba et al. (2006). Second, we have to be careful to control for the date of the vote, as most of the

Table 2.1 RCVs on whole legislative proposals in the last reading before final adoption

Note: Number of proposals adopted in different time periods (according to the date of the EP reading) was deduced from the Commission’s online database PreLex (http://ec.europa.eu/prelex)

RCVs occurred in the last year of the sixth EP and might therefore drive the results. Third, future research on final votes from 2009 onwards will not have to face the problem of a selection bias in RCVs since all these votes are now taken by roll call.

However, at least for votes taken before 2008, we have to consider that the sample might be biased. Luckily, the first problem outlined in the literature, namely the underrepresentation of legislative votes in roll calls, does not need to concern us. As we only study Codecision legislation, non-legislative votes are automatically excluded from the analysis and we do not mix different types of votes. The second problem has to be taken more seriously. Carrubba et al. (2006) discovered that some issue areas are overrepresented in RCVs while others are underrepresented. If this is also the case for the dataset used in the present study, it does not necessarily bias the analysis if voting unity and the factors that influence it do not differ systematically between policy areas. However, if some issues are extremely under- or overrepresented in the data, it would still be good to know. To explore that, we distinguish between different time periods: the whole period under review (Jul 1999-Jun 2009), the period Jul 1999-Jun 2008, when RCVs were only a small proportion of all votes, and the period Jul 2008-Jun 2009, when over 90 % of all votes were taken by roll call. Table 2.2 displays the number of RCVs in the data for each issue area. As the period Jul 2008-Jun 2009 includes nearly a full sample, the percentages presented in the last row should be rather close to the true proportion of votes for each area. Problems might only arise for the period Jul 1999-Jun 2008. However, as can be seen from Table 2.2, differences between the two periods concerning the percentages of RCVs in each area are small. Notwithstanding the small sample size for Jul 1999-Jun 2008, one can thus conclude that issue areas are in general neither overrepresented nor underrepresented in the sample. This is in contrast to the findings of Carrubba et al. (2006). However, the huge differences in the occurrence of RCVs in different issue areas found by them could be due to the fact that they mixed legislative and non-legislative votes. Non-legislative proposals are concentrated in some issue areas. If non-legislative proposals are overrepresented in RCVs, the areas in which they occur most often might be overrepresented as well. This does not apply to our data as it incorporates only legislative proposals. As a result, no particular issue areas are overrepresented in the data used for this study.

Table 2.2 RCVs by issue area for different periods (whole period: Jul 1999-Jun 2009; period with a small number of RCVs: Jul 1999-Jun 2008; period when over 90 % of all votes were taken by roll call: Jul 2008-Jun 2009)

Note: Calculations based on own dataset

The third problem concerning RCVs that has been pointed out in the literature is the risk of overestimating the cohesion of EPGs. This could affect the present study in two ways: On the one hand, Carrubba et al. (2006) might be correct. In particular, MEPs may be more likely to defect from their group to follow the position of their national party in non- RCVs than in RCVs (Trumm 2015). This would mean that EPGs use roll calls strategically to enforce cohesion if they fear that some of their MEPs have reason to diverge, for example, because their minister holds a different position from the one of their group. As a result, research based on RCVs might underestimate voting unity between the MEPs and the minister (because RCVs are called in situations when the EPG forces the MEPs to vote with the group line instead of with their minister). On the other hand, Thiem (2006) might be right. According to her, EPGs only call RCVs if they presume widespread consensus among their members, for example, because none of the component national parties see any reason to diverge (meaning that the ministers of these parties also share the position of the EPG). Then, research based on RCVs might overestimate voting unity between the MEPs and the minister (because issues that are subject to RCVs are less contentious than others).

Yet how can we test whether EPG cohesion is in fact overestimated (and voting unity thus either under or overestimated)? As it is impossible to trace how MEPs voted in non-RCVs and whether MEPs change their voting behavior due to the fact that a vote is recorded, the “true” cohesion of EPGs in non-recorded votes will never be known. However, with the nearly full sample of votes between Jul 2008 and Jun 2009 at hand, we can at least assess whether there were any differences in group cohesion between this last period, when nearly all votes were taken by roll call, and the previous periods. To do so, we calculate the average Agreement-Index (AI) for each of the four main EPGs (as well as for the EP as a whole) for different time periods (see Table 2.3). This AI takes the value 1 if a group is totally cohesive and the value 0 if it is totally divided.⁸ As can be observed, in-group cohesion in the sixth EP (Jul 2004-Jun 2009) is generally higher than in the fifth EP (Jul 1999-Jun 2004).

Furthermore, EPG cohesion does not decline significantly for the period after July 2008 when nearly all votes became RCVs (last column of Table 2.3). On the contrary, for some groups, cohesion is even higher than in the years before. This finding challenges the assumption that cohesion is overestimated in RCVs. If cohesion was in fact overestimated due to a sampling bias, we would expect for the observed levels of cohesion to decline as the sample’s share in the whole population of votes increases. This is not the case, indicating that there is no sampling bias in respect to EPG cohesion in the dataset. Besides, the finding corroborates an assumption made by Hix et al. (2007, 30): Before 2008, only “interesting” votes—when at least one party group took a special interest in the

Table 2.3 Average Agreement-Index (AI) for each of the four main EPGs for different time periods

Note: AI = 1 if an EPG is totally cohesive, and = 0 if EPG is totally divided, calculations based on all final legislative votes between 1999 and 2009

outcome of the vote—became RCVs. When the ID group started requesting RCVs for nearly every vote in 2008, the “uninteresting” votes were also recorded. These “uninteresting” votes concerned technical matters or other less salient subjects where there was little dispute because everyone is in favor. Consequently, on average, RCVs after 2008 display much less dispute in the plenary and also less dispute within the EPGs. Therefore, we observe higher cohesion of EPGs the higher the share of RCVs of the total number of votes taken—a finding that is laid out in more detail in Yordanova and Muhlbock (2015).

Taken together, the above stated findings imply that the data used for the present study is not as seriously biased as could be assumed. First, legislative votes cannot be underrepresented as only legislative votes are used. Second, no particular issue area is overrepresented. Third, EPG cohesion does not seem to be overestimated. Nonetheless, we cannot offer final proof that the data is not biased at all. We cannot rule out that MEPs might alter their behavior if a vote is held by roll call as opposed to normal votes by show of hands. Thus, the results presented here might still not apply for votes that are not recoded. Yet, the behavior of MEPs in RCVs is interesting in itself (Hix et al. 2007, 91). RCVs show how MEPs vote when their voting decision is made public (the records of each RCV are published on the EP’s website and on http://www.votewatch.eu). Not only the EPGs, but the media, interest groups, the general public, and, most importantly for the analysis presented here, national party leaders can also observe the behavior of the MEPs in these votes. If we consider EPGs and national parties as competing principals of MEPs, this means that MEPs might alter their voting behavior in RCVs not only to comply with their EPG but also to please their national party at home. Of course, this poses an additional dilemma: How would MEPs vote in non-RCVs—that is, when their national party cannot survey their behavior? Again, it is impossible to answer this question. Therefore, we have to narrow down the question: How do MEPs vote in RCVs (when at least in theory their national party back home is able to monitor their behavior)? For the research presented in this book the answer to this question is even more interesting than some hypothetical “what would have happened if” considerations, especially due to the fact that—as mentioned above— since 2009 all final legislative votes are taken by roll call and thus the question of how MEPs would have voted in non-roll-called final votes will become obsolete in future analyses.

• [1] Legislative votes might be underrepresented (2) Some policy areas might be overrepresented (3) EPG cohesion might be overestimated