Introduction: What Drives Voting in Legislatures?

Voting in legislatures is a central process in representative democracies. Parliamentary or congressional votes enable politicians to express their preferences on legislation and other issues as well as demonstrate their support for (or opposition to) the government. The standard spatial model assumes that legislators’ and parties’ policy positions are the main drivers of their legislative voting behavior and that, as a result, the median legislator usually decides (under simple majority rule) whether a majority forms to the right or the left of her position. Some recent models assume a role for government-opposition dynamics in legislative voting, for example with governing parties able to restrict the agenda or with opposition parties able to pre-commit to vote against government-sponsored bills (Cox and McCubbins Reference Cox and McCubbins2005; Dewan and Spirling Reference Dewan and Spirling2011). Nevertheless, it is usually assumed that such government-opposition dynamics are only likely to be the norm under specific institutional contexts, for example where a single party controls a legislative majority in a parliamentary system.

In this article we analyze legislative voting in various parliaments from different institutional contexts. In particular we look at whether politicians vote mainly on the basis of their personal or party (left-right) policy preferences or mainly along government-opposition lines. We first apply a scaling method to map the “revealed positions” of parties and individual members of parliament (MPs) in each parliament. We then undertake a vote-by-vote analysis to explore whether left-right policy positions or government-opposition dynamics (or some combination of the two) explain party behavior in roll-call votes. We refer to the outputs of our scaling models as “revealed positions” rather than “ideal points,” as we assume that how legislators vote in roll calls is influenced as much by institutional factors as by their personal ideological preferences. Whereas the scaling analysis is at the individual level, the vote-by-vote regression analysis is at the party level. We hence see these two analyses as complementary, as a way of cross-checking the empirical regularities identified in each of the analyses.

Our primary aim is not to directly test a particular theoretical model or set of theoretical propositions, but rather to identify some empirical regularities that raise questions about existing and future theoretical work on legislative behavior. For example, what are the main dimensions of political conflict in modern democratic legislatures? How do the (left-right) policy preferences of legislators and parties influence how they vote? Is legislative conflict primarily between members of governing parties and members of opposition parties, irrespective of their policy preferences? And how does the institutional design of a democracy—such as whether a country has a presidential or a parliamentary regime, or whether there is a single-party, coalition or minority government—influence the pattern of parliamentary voting?

Since the late 1990s there has been an explosion of empirical research on legislative voting. The impetus came from new scaling methods (for example, Poole and Rosenthal Reference Poole and Rosenthal1997; Poole Reference Poole2005). Armed with these new technologies, the increased availability of roll-call data on the internet from other legislatures and the development of computer power, scholars began to study voting in legislatures outside the United States. Although some scholars (for example, Morgenstern Reference Morgenstern2004; Hansen Reference Hansen2006a, Reference Hansen2006b; Carey Reference Carey2007a; Cox and McCubbins Reference Cox and McCubbins2011) look at roll-call voting in several parliaments, most of this new research has examined single parliaments. Our goal is to fill this gap by producing spatial maps of several legislatures using the same information (roll-call votes) and broadly the same time period, and using the same methodology and analysis.

We look at 16 legislatures: Australia, Belgium, Brazil, Canada, Chile, Czech Republic, the European Parliament, France, Israel, Mexico, New Zealand, Peru, Poland, South Korea, United Kingdom and the United States. We use these cases for two reasons. First, roll-call voting data are publicly available only for a limited number of legislatures, and there is considerable variance in the quality and quantity of available data in many cases. In addition to the data in the public domain, we have collected new data from the Czech Republic, Belgium and the European Parliament. Despite some differences in quality and quantity, the data from these legislatures are reliable and include a sufficient number of votes to allow us to estimate legislators’ positions using a standard scaling metric. Second, these legislatures include cases of all possible combinations of two key institutional variables: regime type (parliamentary or presidential) and form of government (single-party majority, multi-party majority, and minority). Of course, other institutional variables, such as the electoral system, also shape the number of parties and voting behavior inside parliaments (for example, Carey and Shugart Reference Carey and Shugart1995; Hix Reference Hix2004; Shomer Reference Shomer2009). However, in this article we only focus on regime type and form of government, and postpone the relevance of other important factors to future research.

Our main empirical finding is that the standard spatial (median-voter) model fits legislative voting in particular institutional contexts: in presidential systems with coalition governments or parliamentary systems with minority governments, where coalitions have to be built issue by issue. In all other institutional contexts, government-opposition splits dominate policy-based voting by parties or legislators. These results raise some questions about how theories of legislative politics might be adapted to better fit empirical regularities. We also find that the existing scaling methods can be used to estimate the revealed positions of legislators across the democratic world, and that two dimensions are sufficient to explain a large fraction of the variance in voting behavior in most parliaments.

The rest of the article is organized as follows. In the next section we discuss some of the existing theories of legislative voting. We then investigate the dimensionality of voting in our 16 cases by looking at the “voting maps” produced by a standard geometric scaling technique. We subsequently look at how much of the voting in each parliament can be explained by the government-opposition status of an MP as opposed to his/her left-right policy position. We also consider the case of the Czech Republic, where the shift between minority and majority status in successive parliaments led to a significant change in the structure of voting.

Theories of Legislative Politics

Most existing theories of legislative politics, as well as the scaling methods for testing these theories, were developed in the context of the US Congress. Applying these theories beyond the US Congress must be done with great caution. Nonetheless, the existing theories do provide some useful benchmarks.

Most current theoretical models of legislative politics start with the standard assumptions of spatial voting in the legislative context (cf. Black Reference Black1948, Reference Black1958; Romer and Rosenthal Reference Romer and Rosenthal1978; Shepsle Reference Shepsle1979; Shepsle and Weingast Reference Shepsle and Weingast1981; Tsebelis Reference Tsebelis2002; Poole Reference Poole2005). One such model is the median-voter model with randomly assigned proposal power, which is sometimes called the “floor agenda model” (cf. Romer and Rosenthal Reference Romer and Rosenthal1978; Krehbiel Reference Krehbiel1998; Cox and McCubbins Reference Cox and McCubbins2005, 38–41). In this model, one of the legislators makes a proposal and the others are free to propose amendments. Amendments are voted on one by one and then a final vote is taken on the amended bill, all by a simple majority vote. In any vote, each legislator (with single-peaked preferences) has a binary choice: between a policy represented by a bill and the status quo reversion point. If the distance between a legislator’s ideal point and the bill is smaller than the distance between the legislator’s ideal point and the status quo, the legislator votes for the bill; otherwise she votes for the status quo.

A crucial assumption of this model is that legislators are not constrained by parties. They may belong to parties, but parties are not able to enforce cohesion in votes. Instead, each legislator votes for the option (status quo or proposed bill) that is closest to her ideal point. Members of the same party might vote together, but this is because their ideal points are on the same side of a cut-point in a vote and not because the party has forced them to do so (Krehbiel Reference Krehbiel1993).

The floor agenda model predicts that final votes, in a one-dimensional policy space, will come down to a choice between the location of the status quo and the policy proposed by the median floor member. Hence, cut-points should be located at the mid-point between the legislator who prefers the position of the median floor member to the status quo by the smallest margin and the legislator who prefers the status quo to the position of the median floor member by the smallest margin. If status quos are evenly distributed in a policy space, cut-points should also be evenly distributed, which would then allow the “ideal points” of legislators on an underlying policy dimension to be identified from roll-call votes (Poole Reference Poole2005).

A variation of the standard floor agenda model is the “cartel agenda model,” as proposed by Gary Cox and Mathew McCubbins (2005). In this model a particular legislator or group of legislators can restrict the agenda. For example, in the US Congress the majority party can decide which proposals are put to the floor of the House or Senate. Similarly, in parliamentary systems, if the party or parties in government control a majority of parliamentary seats, then the government has a de facto monopoly on agenda control (Döring Reference Döring2001).

Adding agenda control restricts the set of viable policy outcomes. This is because the median member of the agenda cartel will prevent an issue from being placed on the agenda if the outcome on the floor will be further from her most-preferred policy than the status quo. Specifically, there should be no proposals on any issue where the status quo is closer to the position of the median member of the floor than to the position of the median member of the agenda cartel.Footnote ¹ Nevertheless, the agenda cartel will still allow votes to be taken on the plenary floor if the distance between the median member of the legislature and the position of the agenda cartel is smaller than the distance between the status quo and the position of the agenda cartel. As a result, some legislative votes will divide the members of the agenda cartel (for example, the governing parties) against the members of the non-agenda cartel (for example, the opposition parties), while others will still split along left-right lines. Hence, the cartel agenda model predicts a mix of majority-bloc (government) versus minority-bloc (opposition) voting and policy-based (left-right) voting.

Finally, several other theoretical ideas underpin what can be called a “Westminster model” of legislative politics (Cf. Lijphart Reference Lijphart1999). On one side is the governing party, which can enforce party cohesion in votes via a variety of carrots and sticks. As carrots, the governing party can offer promotion to ministerial office or key committee positions (Benedetto and Hix Reference Benedetto and Hix2007; Kam Reference Kam2009). As a stick, the main weapon is the threat of a vote of confidence (Huber Reference Huber1996; Diermeier and Feddersen Reference Diermeier and Feddersen1998). Faced with the possibility of cabinet resignation and/or an early election, a governing party “backbencher” will vote for a government proposal even when the policy on offer is further from her ideal point than the status quo.

On the opposition side, Dewan and Spirling add an assumption that opposition legislators can credibly pre-commit to oppose the government in most legislative votes (Dewan and Spirling Reference Dewan and Spirling2011). Although members of each party have similar policy preferences on most issues, there is sufficient heterogeneity within the governing and opposition members such that on some policies, members of the opposition find themselves closer to the government than to the opposition leadership. With sincere voting, some members of the opposition are hence likely to prefer a proposal of the government (the agenda setter) to the status quo. However, if the opposition members can pre-commit to vote cohesively against a government proposal, then the government will be forced to make a proposal closer to the status quo than would be the case with sincere voting. This is because with all the opposition members voting to keep the status quo, the de facto pivotal member is closer to the status quo than the median floor member under sincere voting. In all Westminster-style systems, some votes, such as “non-whipped” votes, will not follow this strict government-opposition divide. Nevertheless, the Westminster model, as we have characterized it here, predicts that most cutting lines will split governing members against opposition members, and not between individual members or parties along a left-right policy dimension (cf. Laver Reference Laver2006; Spirling and McLean Reference Spirling and McLean2007).

There are numerous other models of legislative politics (for example, Baron and Ferejohn Reference Baron and Ferejohn1989). However, the models we describe here are particularly useful for our purpose because they make specific and different assumptions about the institutional context of legislative politics, as Figure 1 shows.Footnote ²

Fig. 1

Institutional contexts of our cases Note: France and Poland have semi-presidential regimes. However, because the government (via the president) can dissolve the parliament in both systems, and hence enforce party discipline, we group these cases together with the parliamentary systems. We categorize the European Union (EU) as a separated-powers system because the EU Commission has a monopoly of legislative initiative, the commission cannot dissolve the European Parliament and the European Parliament can only remove the commission by a special oversized majority vote (Hix and Høyland Reference Hix and Høyland2011). WM=Westminster model, CAM=cartel agenda model, FAM=floor agenda model.³

First, whereas the Westminster model assumes that parties can enforce legislative cohesion, the floor agenda model and the cartel agenda model assume that parties cannot. This difference fits the difference between parliamentary and presidential systems (cf. Cheibub Reference Cheibub2007; Carey Reference Carey2007a, 116–35; Samuels and Shugart Reference Samuels and Shugart2010). In parliamentary systems, governing and opposition leaders can offer promotion to cabinet or opposition leadership positions, respectively. Also, the possibility of parliamentary dissolution is a threat against legislative rebellion on the government side and an incentive to vote collectively to try to defeat the government on the opposition side. In contrast, in presidential systems, where there is independent election of the chief executive and the legislature,Footnote a president might be able to promote some people to cabinet positions, but a defeat of the governing party by the opposition on a particular bill does not threaten the survival of the executive. There is some empirical support for this argument. For example, using data from 19 systems, Carey finds that on average, governing parties in parliamentary systems are 17 percent more cohesive than governing parties in presidential systems (Carey Reference Carey2007a, 104).Footnote ⁴

Second, whereas the Westminster and cartel agenda models assume that governing parties can restrict the legislative agenda, the floor agenda model assumes no agenda restrictions. These assumptions relate to whether a government has majority control of the legislature and whether a government is a single-party or a coalition government. Where a single party controls the agenda and commands a majority in a legislature, the leader of that party is effectively a legislative dictator. This is most clearly the situation in a single-party government in a parliamentary system. As one backbench member of the British House of Commons once put it: “we are reduced to throwing paper airplanes at the government bulldozer”.Footnote ⁵ In a presidential system, in contrast, where one party controls a majority of seats in the legislature and sets the agenda, the party can use this agenda-control power to restrict the legislative agenda. However, because of the separation of executive and legislative elections, the majority party will have few powers to enforce cohesion.

Contrast this monopoly of agenda-setting power in a single actor with the sharing of agenda setting between several actors, as is often the case in presidential regimes (where the executive and legislature both have agenda-setting power) as well as under coalition governments in both parliamentary and presidential systems (cf. Cheibub Reference Cheibub2007). In these contexts it is more difficult for a single actor to restrict the set of issues that comes to the legislative floor. In a coalition government, different parties will have different ministerial portfolios, which can lead to different actors proposing to move policies in different directions (Laver and Shepsle Reference Laver and Shepsle1996; Tsebelis Reference Tsebelis1999; Strøm, Muller and Bergman Reference Strøm, Muller and Bergman2011). Related to this, in legislatures with coalition governments, legislative committee chairs are usually shared between several parties that each have access to the agenda (Mattson and Strøm Reference Mattson and Strøm1995; Döring Reference Döring2001). As a result, a wider range of issues is likely to be voted on when there is a coalition government than when there is a single-party government. Nonetheless, majority coalitions in parliamentary regimes are more likely to be able to enforce party discipline than majority coalitions in presidential regimes (Lupia and Strøm Reference Lupia and Strøm1995; Diermeier and Feddersen Reference Diermeier and Feddersen1998). This suggests that government-opposition dynamics should be stronger in our cases of majority coalitions in parliamentary regimes than in our cases of majority coalitions in presidential regimes.

Finally, minority governments, regardless of their regime type, are forced to build legislative coalitions issue by issue (Strøm Reference Strøm1990; Tsebelis Reference Tsebelis2002, 97–9). This is not to say that government-opposition conflicts do not exist under minority government, but rather that minority governments cannot adopt policy without the help of some opposition legislators. In practice, this suggests that under minority governments, on any given policy issue the median legislator or party is pivotal for both government bills and bills or amendments proposed by opposition members. As a result, in this context legislative voting is likely to be dominated by parties’/legislators’ (left-right) policy positions rather than by government-opposition splits. Nevertheless, because of the ability of parties to discipline their members in parliamentary systems as compared to presidential systems, there is again likely to be a difference in how this works across the two types of regimes: with parties as the units of analysis under minority government in parliamentary systems, and individual legislators as the units of analysis under minority government in presidential systems.

In sum, the main theories of legislative politics suggest the following: (1) that left-right policy positions should be more correlated with roll-call voting than government-opposition dynamics, where there are minority governments (in either parliamentary or presidential regimes) and where there are majority coalition governments in a presidential regime (the contexts that fit the floor agenda model); (2) that government-opposition dynamics should dominate voting where there are majority single-party governments in a parliamentary regime (under the Westminster model) and (3) that both government-opposition voting and left-right voting should exist where there are majority coalition governments in a parliamentary regime and majority single-party governments in a presidential regime (the contexts that most closely fit the cartel agenda model).

Estimation of the Voting Space in 16 Legislatures

To investigate these propositions we apply a standard scaling metric to the roll-call votes in a range of parliaments: Clinton, Jackman and Rivers’ IDEAL. IDEAL is a leading Markov Chain Monte Carlo-based method of ideal point estimation. Like other scaling methods, IDEAL provides a good way of data reduction, and produces model based estimates. IDEAL, like other scaling methods, needs identification restrictions, such as determining the polarity of each dimension, normalization and rotation of dimensions. We fix the polarity of each dimension by constraining the location of a known left-wing (right-wing) legislator or political party to the negative (positive) value on a dimension. We also rescaled the estimated ideal points such that they all lie within a unit circle. Finally, we rotated the estimated ideal points so that the first dimension is the one associated with the largest variance (or eigenvalue). Using those identification restrictions for our 16 cases, we obtained legislator maps that are directly comparable.

As a robustness check we also use Poole and Rosenthal’s NOMINATE (Poole and Rosenthal Reference Poole and Rosenthal1997; Clinton, Jackman and Rivers Reference Clinton, Jackman and Rivers2004). NOMINATE has become an industry standard for estimating legislators’ revealed preferences from roll-call votes because it is relatively easy to apply, and requires limited computer capacity to produce estimates on several dimensions. Although each method has its advantages and disadvantages, the results we obtained are highly correlated and largely comparable.Footnote ⁶ Given that the results from the two methods are highly correlated, we report only the IDEAL results.Footnote ⁷

Roll-call votes have been collected and analyzed in published and publicly available research for over 30 legislatures.Footnote ⁸ However, raw roll-call data are only in the public domain for a limited number of legislatures, and for many of these cases the data are not suitable for applying a scaling metric, as either there is an insufficient number of votes (as with the data from many of the Latin American parliaments and Japan) or the votes are heavily lopsided (as in the Russian case).

There is, of course, considerable variance in the meaning of the roll-call votes in legislatures for which good-quality roll-call data exists, and hence questions about their comparability. Some legislatures, such as the Czech Republic and the European Parliament, have over 2,000 votes in a session while others, such as Peru or France, have fewer than 200 votes in a similar period. There is also a potential problem of selection bias: where roll-call votes are strategically requested by parties to enforce certain types of behavior (for example, Hug Reference Hug2010). The subject of votes varies across parliaments, some legislatures have more votes on budgetary issues than others, and some periods are dominated by particular nationally specific issues. The number of roll-call votes also varies by country, which is in part related to the rules governing what issues get to a vote (Carey Reference Carey2009). And the legislative rules of procedure of course vary considerably across countries. All in all, one might think that because so many such factors co-vary by country that little systematic can be identified from studying roll-call voting across countries.

Nevertheless, the data at hand cover real-world legislative votes on important public policies such as institutional reforms, foreign policies, fiscal policies and budget. Also, because we scale each legislature separately,Footnote ⁹ we do not need to weight the analysis by the number of roll-call votes in each legislature. Moreover, the sample size in each case is not correlated with the frequency of government-opposition splits. And, conceptually, it is not clear why selection effects (or selective recording of votes) should lead to more government-opposition voting. We also show that the standard scaling methods can be applied to produce, at least to some extent, comparable ideological maps in different legislatures. Furthermore, we illustrate that our findings hold when looking at over-time variation in the Czech Republic. By focusing on over-time variation in this case, we try to isolate the effect of a change in agenda control (between majority and minority governments) while keeping constant the underlying regime structure, party system, legislative rules, strategic incentives and perhaps even issues.

For the legislatures that have good quality roll-call data available, we proceeded as follows. We first chose one full term of the legislature in either the late 1990s or early 2000s, or a part of term if the data from a full term were not available (in the Israeli case).Footnote ¹⁰ We then applied the same criteria to decide which votes and parliamentarians to exclude from the analysis: we dropped all lopsided votes (in which less than 10 percent of legislators were on the minority side) and all legislators who voted fewer than 25 times. We then applied the same optimization algorithm: the latest versions of IDEAL and W-NOMINATE. We estimated two-dimensional models for each parliament for two reasons. First, dimensions higher than the first two rarely have a clear interpretation. Second, the increase in the goodness-of-fit statistics, such as the correct classification scores, is negligible for higher dimensions. In addition, estimating a two-dimensional policy space is standard practice in the literature (for example, Poole and Rosenthal Reference Poole and Rosenthal1997; Londregan Reference Londregan2000; Morgenstern Reference Morgenstern2004).Footnote ¹¹

Table 1 summarizes the datasets and the goodness-of-fit statistics associated with our cases. The first noteworthy finding is that the scaling results in most legislatures indicate a predominantly one-dimensional policy space. Nevertheless, there is some variation between the cases. For example, the second dimension explains a significant amount of variance in the Canadian, Israeli, Polish and South Korean legislatures. Eigenvalues reported in Table 1 also suggest a third dimension in the Czech legislature, and to a lesser extent in the Polish and Israeli legislatures and in the European Parliament, but not in any of the other cases. At the other extreme, the Australian, French and New Zealand legislatures look essentially one-dimensional, as the second and higher dimensions explain very little in these cases.

Table 1

Dimensionality in Parliaments: Goodness-of-Fit Statistics

Note: *This number slightly decreased, indicating that the second dimension had no explanatory power. APRE=aggregate proportional reduction in error, where the classification errors of the model are compared to those of the naïve benchmark of assuming that all MPs vote with the majority.

Figures 2–5 show the “voting maps” of the two-dimensional estimates produced by IDEAL for each legislature.Footnote ¹² In each map, the distance between any two legislators illustrates how often they voted the same way in the roll-call votes in a given period. So, if any two legislators voted the same way in every vote, they would be located in exactly the same place, while if they voted on opposite sides in every vote, they would be located on opposite sides somewhere around the rim of the unit circle in the figure. A note of caution, however: because each legislature is scaled independently, distances in one figure cannot be compared to distances in another figure. Hence, within a legislature, if one party’s members are close together while another’s members are dispersed, we can infer that the former party voted more cohesively than the latter. However, across two legislatures, if a party’s members in one legislature are close together while another party’s members in another legislature are more dispersed, we cannot infer that the former is more cohesive than the latter, as there may simply have been more interparty vote splits in the former than in the latter legislature.

Fig. 2

Parliamentary regimes with single-party governments

Fig. 3

Parliamentary regimes with (majority) coalition governments

Fig. 4

Presidential (separated-powers) regimes with single-party governments

Fig. 5

Presidential (separated-powers) regimes with coalition governments

Nevertheless, the maps do provide suggestive evidence of whether voting within a legislature is mainly driven by government-opposition splits or by parties’ left-right policy positions. In each voting map, the members of the governing party are indicated by an ellipse, and a possible left-right dimension is indicated by a dotted line from the most left-wing major party to the most right-wing major party (the names of the parties and their political affinities are listed in the Appendix).

Let us stress again what we mean by government-opposition voting. When legislators vote based on whether they belong to a governing or opposition party, we say that voting takes place along a government-opposition “dimension.” That is, government-opposition dynamics are at work when a majority of members belonging to government parties votes against a majority of opposition parties’ members. By government-opposition dimension, we do not mean a continuous policy dimension that is always independent of a left-right policy-based dimension. In many cases, the two dimensions are highly correlated, as coalitions form based on the left-right ideologies of the parties. Only when the two dimensions are uncorrelated can we investigate whether voting is shaped by left-right policy preferences rather than by government-opposition dynamics.

Figure 2 shows the four cases of single-party majority government in parliamentary regimes. The first dimension in all these maps clearly captures the government-opposition divide. With the two-party systems in the 1996–98 Australian House of Representatives and the 1990–93 New Zealand House of Representatives, it is impossible to identify government-opposition voting independently from the left-right policy preferences of the parties. In contrast, in the 1994–97 Canadian House of Commons and the 1997–2001 British House of Commons, while the first dimension is clearly a government-opposition split, the second dimension appears to capture left-right divisions within the opposition group of parties.

The maps of the four cases of coalition governments in parliamentary regimes, in Figure 3, are more varied. Voting in the 2003–07 Belgian Federal Chamber of Representatives was similar to the single-party government cases, in that the legislators from the parties in the governing coalition are on the left while the legislators from the opposition parties are on the right, and the second dimension captures left-right divisions within both the opposition bloc and the governing coalition. Within each bloc, on the second dimension we also observe some conflicts between the Dutch-speaking parties (on the upper part) and the French-speaking parties (on the lower part). In the 1997–2002 French National Assembly and the 1999 Israeli Knesset, the government-opposition split correlates with parties’ left-right positions. The meaning of the second dimension is less clear in these cases. In Israel, for example, the location of the parties on the second dimension within the government and opposition blocs suggests that the second dimension captured several issues, such as religious parties versus secular parties. Meanwhile, in the 1993–97 Polish Sejm, there was variation within the government and opposition blocs on both dimensions, and while the more left-wing party in the opposition is lower on the second dimension, the more left-wing party on the governing side is higher on the second dimension.

Turning to the presidential regimes, Figure 4 shows the three cases of single-party majority government (in which one party holds the presidency, all the cabinet seats and a majority in the legislature) in these systems, along with the one case of minority government in a presidential regime (where a party that holds the presidency does not command a majority in the legislature) (for example, Cheibub Reference Cheibub2007). Again, there are a variety of patterns. In both the 1999–2000 Peruvian Congress and the 2004–07 South Korean National Assembly, the first dimensions seems to be mainly related to the government-opposition dimension, while the second dimension captures left-right policy differences between the parties in opposition. Meanwhile, the two-party system in the United States means that government-opposition voting and the left-right policy preferences of the parties cannot be separated for the US House of Representatives. In contrast, in the 2003–06 Mexican Chamber of Deputies the first dimension captures left-right policy preferences and government-opposition dynamics, with the conservative National Action Party (PAN) in government on the right, but with a minority of seats, the centrist Institutional Revolutionary Party (PRI) in the center, and the left-wing Party of the Democratic Revolution on the left. The relative positions of the parties suggests that the minority PAN government relied more on the party closest to it on the left-right dimension (the PRI) to build majority coalitions in the legislature.

Figure 5 shows the three cases of coalition governments in presidential/separated-powers regimes and the one case of minority government in a parliamentary regime. In all four of these cases, the first dimension appears to be correlated with the left-right positions of parties. In the 1995–98 Brazilian Chamber of Deputies the government-opposition camps are identifiable, but there was some variance within the government and opposition blocs. Similarly, in the 2004–09 European Parliament, the political groups are lined up from left to right on the first dimension, while the political groups that have members in the European Union executive (the European Commission) are close together on the second dimension but divided on the first dimension. In contrast, the 1998–2000 Chilean Chamber of Deputies looks similar to some of the voting patterns under coalition governments in parliamentary systems, with a clear divide along government-opposition lines.

Finally, voting in the 1998–2001 Czech Chamber of Deputies was clearly along party lines, with cohesive party-line voting and the parties lined up from left to right on the first dimension. The positions of the parties on the map also show that the minority governing party (CSSD) built coalitions issue by issue, either with the former communists (KSCM), or sometimes with the Christian democrats (KDU) and liberals (US), and less often with the conservatives (ODS), who are far from the governing party on both dimensions.

Estimating the Relative Significance of Left-Right and Government-Opposition Voting

These maps consequently allow some inferences to be drawn about the substantive meaning of the revealed dimensions of voting in these 16 legislatures. However, these inferences rely on a rather ad hoc understanding of the left-right positions of the parties. We consequently try to interpret the relative significance of left-right and government-opposition voting in a more systematic way. One estimation strategy would be to regress each dimension of IDEAL as a function of left-right party positions and government-opposition variables.Footnote ¹³ An important shortcoming of such an exercise would be that pooling data from different countries cannot be logically justified without strong assumptions. As a result, we follow an approach that does not require pooling votes from different countries. We thus look at every vote in each legislature and estimate how much of the behavior of each individual legislator in each vote can be explained by his/her government-opposition status or left-right policy position (cf. Mian, Sufi and Trebbi 2012).

For each of the ~11,000 votes we estimate the following cross-sectional specification:

$$Vote_{i}^{{r,J}} =\alpha ^{{r,J}} {\plus}\beta ^{{r,J}} \,LR^{J} {\plus}\gamma ^{{r,J}} GO^{J} {\plus}{\varepsilon}_{i}^{{r,J}} ,$$

where i indicates a member of legislature J and r refers to a specific roll-call vote in that legislature. LR ^J is the left-right position of legislator i’s party in parliament J, while GO ^J is a dummy variable indicating whether the legislator’s party in parliament J is in government (coded 1) or opposition (coded 0). Finally, ${\varepsilon}_{i}^{{r,J}} $ is the error term. We estimate the model using the linear probability model with heteroskedasticity-robust standard errors. Given that legislators belonging to a given party vote similarly, we cluster the observations by political party.Footnote ¹⁴

Ideally, we would like to have exogenous measures of individual legislators’ left-right policy preferences rather than aggregated at the level of political parties. However, except for a few cases—such as the US Congress or the European Parliament—individual-level measures either do not exist or are unreliable (Laver Reference Laver2006). In contrast, there are reliable exogenous measures of party positions on a range of policy issues, and previous research has found that these measures correlate with party voting behavior inside legislatures (for example, Hix, Noury and Roland Reference Hix, Noury and Roland2006; Zucco and Power Reference Zucco and Power2009). So, for the left-right location of each party, we use the “expert estimates” of parties’ left-right positions in the Benoit-Laver and Wiesehomeier-Benoit datasets (Benoit and Laver Reference Benoit and Laver2006; Wiesehomeier and Benoit Reference Wiesehomeier and Benoit2007).Footnote ¹⁵

Our approach will not provide meaningful results if the left-right and government-opposition variables are highly correlated. Given that our explanatory variables are party-level data, and do not vary within a given party, perfect multicollinearity is an issue in some cases. In particular, in two-party legislatures (United States and New Zealand) we have perfect multicollinearity. In addition, in a few other cases (Australia, Chile, France, Peru and, to a lesser extent, Brazil and Mexico) we observed a high correlation between left-right and government-opposition.

Another limit of our approach is a potential measurement error in our left-right variable. Measurement errors do not always affect the properties of ordinary least squares (OLS). But if potential measurement errors are correlated with our left-right variable, then our estimates will suffer from an attenuation bias and the other variables will be estimated with some bias as a result. To address this potential problem we use Carroll et al.’s (Reference Carroll, Ruppert, Stefanski and Crainiceanu2006) error correction method, known as simulation extrapolation (Simex) (cf. Benoit, Laver and Mikhaylov Reference Benoit, Laver and Mikhaylov2009). For each vote we look at whether each explanatory variable is significant at 5 percent. For each legislature we report the number and proportion of cases in which the left-right dimension is significant, and compare it to the number and proportion of cases in which the government-opposition dimension is significant.

As indicated by Mian, Sufi and Trebbi (Reference Mian, Sufi and Trebbi2012), the approach we use has some useful features and some evident drawbacks. One useful feature is that we are able to directly aggregate a large number of heterogeneous votes. In addition, by focusing on statistical significance as opposed to the direction and size of coefficients, we can abstract from arbitrarily classifying roll-call votes as pro-government or in a right-wing direction. A disadvantage, though, is that all the important legislative information contained in specific bills and amendments is lost. In particular, votes on important policy issues are treated identically to votes on trivial issues.

The results of our vote-by-vote regressions are reported in Table 2. They indicate the relative importance of our left-right variable versus the government-opposition variable. For Canada, for example, the government-opposition split was significant in 450 out of 498 regressions, whereas the left-right dimension was significant in 299 regressions. We therefore infer that in Canada during this period the government-opposition split was a more important driver of legislative voting than left-right party policy positions. In only one case, one variable explains the voting behavior: in Belgium, where votes are overwhelmingly explained by government-opposition status. In all other cases, both parties’ left-right policy positions and the government-opposition status of a legislator’s party are important.

Table 2

Left-Right and Government-Opposition Determinants of Legislative Voting

Note: The method used is linear probability (OLS) with clustered standard errors. Australia, New Zealand and the United States are excluded because of almost-perfect collinearity between left-right and government-opposition. The last two columns report the estimates by simulation extrapolation method for error correction.

To check whether measurement errors affect our estimates, we used standard error data reported by Benoit and Laver (2006) to correct for measurement error of the left-right variable. The results reported in Table 2 are qualitatively similar to the estimates produced by simple OLS regression analysis. However, we found different estimates for two cases. After error correction, we find that for Poland, left-right and government-opposition variables are equally important. For Brazil, however, we find that left-right position is now more important than government-opposition status.

Nevertheless, another key finding is that government-opposition status generally trumps left-right policy positions in almost all cases. Only in the European Parliament and the Czech Republic case of minority government do we find that left-right policy positions are stronger predictors of legislative voting than government-opposition status.Footnote ¹⁶ Government-opposition status dominates left-right preferences across all institutional contexts. These finding hence corroborate the inferences from the voting “maps.”

Finally, the findings for the two legislatures in the Czech Republic are particularly interesting. In this country, the minority CSSD government in 1998–2002 was replaced by a majority CSSD-KDU-US coalition government in 2002–06. If the institutional context of legislative voting—whether a majority or a minority government controls the agenda—matters, we should expect a shift in the pattern of voting in the two parliaments. Specifically, we should expect mainly left-right voting during the period of minority government in 1998–2002 and mainly government-opposition voting under the majority coalition government in 2002–06.

This is exactly what we found in the vote-by-vote analysis of the predictors of voting in the two Czech parliaments.Footnote ¹⁷ As Table 2 and Figure 6 show, when the social democrats (CSSD) only controlled a minority of seats in parliament, which meant significant power for the median party and coalitions built issue by issue, legislative votes followed left-right lines more than government-opposition status. Then, after the 2002 election, when the CSSD formed a majority coalition with the Christian democrats (KDH) and the liberals (US), legislative votes followed government-opposition lines more than left-right party positions. The shift from left-right to government-opposition voting between the two parliaments is most clearly illustrated by the dramatic shift in the positions of the KDH and US MPs: from the center-right in the 1998–2002 parliament to the government (on the left) in the 2002–06 parliament.

Fig. 6

Case Study of the Czech Republic