ACCOUNTS OF VOTER BEHAVIOR: CHOICE OVER PLATFORMS

The spatial model (Downs 1957; Hotelling 1929) characterizes party competition in a two-party system where the winner can implement his or her policy platform. Under this system, each voter chooses a party such that the outcome is spatially closest to his or her own bliss point. The result of this model is well known; the two parties will adopt the policy position of the median voter. However, in the Downsian world, a two-party system where the winner can implement his or her preferred policy, policy outcome is identical to the winner's position (assuming a binding platform), and thus the same result can be achieved by either voting over policy outcomes or voting for the party whose position is the most similar to the voter's position. Empirical research adopts the latter motivation—voting over platforms—and formulates voter utility for parties as negatively related to the distances between the voter's and the parties' issue positions (Enelow and Hinich 1984).

With a few exceptions, the vast empirical literature holds that voter evaluation of parties depends on party positions on issues (platforms). This literature revolves around a major point of contention: whether voters prefer parties whose platforms are most similar to their own or parties whose platforms are more extreme than their own. Proponents of Proximity Model argue that voters prefer parties that are ideologically similar to their own views on the issues (e.g., Blais et al. 2001 and Westholm 1997). Disparately, supporters of the Directional Model argue that voters prefer parties that are ideologically in the same direction as, but are more intense than, their own views on the relevant issues (e.g., Macdonald, Listhaug, and Rabinowitz 1991 and Macdonald, Rabinowitz, and Listhaug 2001). Despite the conviction of scholars within each camp, numerous studies comparing the explanatory power of the two models in accounting for voter choice find mixed evidence (e.g., Cho and Endersby 2003, Lewis and King 2000, and Pierce 1997).

In the heated debate between the two camps, Downs's original argument of outcome-oriented voting has been set aside. But as Bailey (2001, 6) reminds us, to the extent that a voter wakes up in sweat in the middle of the night thinking about politics, it is policy outcomes and their immediate effect that are on his or her mind. According to Downs, the voter “weigh(s) the performance that the opposition party would have produced in period t if it had been in power” (40).¹

, where t is the current period (40). While still assuming a winner-take-all setup, Downs extends this decision rule to multiparty systems: “In the latter, the voter follows the same rules as in the former, but compares the incumbent party with whichever of the opposition parties has the highest present performance rating, i.e., would yield him the highest utility income if it were now in office” (47).²

In the following section I present a model in line with the (neglected) Downsian notion of outcome-oriented behavior. Contrary to the voluminous empirical literature, I contend that even when focusing on issues, voter behavior is largely policy-oriented. Under my framework, the rationale for preferring a party whose positions are different from (and not necessarily in the same direction as) the voter's views is policy oriented—voters endorse parties insofar as the parties pull policy outcomes in a desired direction.

VOTING TO ACHIEVE OUTCOMES: A MODEL OF INSTITUTIONAL BALANCING

The Compensational-Vote Model

Under the proximity model, the utility of voter i(i = 1, …, n) for party j(j=1, …, m) is inversely related to the ideological distance between i and j. In one dimension,

where v_i is the ideal point of voter i, p_j is the position of party j, and β₁ is an unknown constant.

As mentioned above, representation or expression of opinions is only one motivation for choosing one party over another. Voters might also use their vote to shift policy outcomes toward their ideal points. According to this logic, they reward parties that pull outcomes in their direction and penalize parties that pull outcomes away from them. How do voters perceive political outcomes? One possibility is that voters utilize a “naïve understanding of democracy” in which policy outcome is a weighted average of policy positions of parties in the legislature, where the weights are the relative impacts of the different parties. Policy outcome is then calculated by

where s_j is the relative impact of party j, such that

and s_j ∈ [0, 1) ∀_j. The relative impact of each party depends on the distribution of power in the parliament, and in particular, on the composition of the parliament, the composition of the governing coalition, and portfolio allocation within the government. While theories of bargaining and coalition formation are not within the scope of this study, I address this issue here briefly.

One might wonder why I allow all parties in the legislature to have an impact greater than zero. After all, some parties are highly unlikely to be members of the government. There are several motivations for not completely discounting the role of small nongovernmental parties. First, in some cases opposition parties are often partners in ad hoc pacts with factions of the governing coalition and thereby get part of their agenda implemented. Second, in some systems (e.g., Norway), the governing coalition is often a minority government, such that it structurally depends on factions of the opposition supporting it (Strøm and Leipart 1993). Third, some parliamentary systems allow the opposition to have partial control over policy formation via authority in permanent legislative committees (including committee chairmanship), conferences between parliamentary leaders, and other mechanisms (Strøm 1990, 207–11). Finally, even uninvolved in the policy process itself, opposition parties affect policy outcomes by shaping the discourse, placing issues on the agenda, and forcing established parties to address these issues. (See Avakumovic 1978 for discussion of the effect of the CCF-NDP in Canada and Meguid 2002 for the impact of rising parties in western Europe.)

When assessing each party, outcome-oriented voters may entertain a counterfactual: how did politics look in the old days when nobody cared about issue X? Utility from policy-motivated voting is then represented by

where β₂ is an unknown constant and

is a counterfactual policy outcome—an outcome that would have been produced had all parties except party j taken part in the policy-formation process:

The intuition behind the bracketed term in Equation (3) is a counterfactual analysis in line with Downs's (1957) counterfactual discussed above (40). If party j pulls the outcome closer to the voter, this term is positive. If j pulls it away from the voter, it is negative. The voter's utility for party j approaches maximum when P approaches the voter's bliss point and

is far from it. Since the model describes an outcome-oriented yet naïve voter, I assume that (from the voter's point of view) in j's absence other parties do not relocate to fill the “vacuum,” nor do their impacts change relative to one another. Note that this formulation does not imply that the voter believes the impact of his or her participation in the elections to be the presence or absence of the party from the map. Rather, he or she is concerned with the impact of the party.

is the marginal impact of party j in policy space, and hence

is the marginal benefit of party j for voter i.

I incorporate the two motivations and normalize β₁+β₂ to 1. In addition, I include a parameter (θ) that indicates the salience of issue voting⁵

In a multidimensional setup with d dimensions, θ is a d×1 vector where each element represents the salience of a certain issue area.

and allow for individual-level background variables such as socioeconomic background to affect the utility for each party in a different fashion:

where β∈(0, 1) is a relative weight of the two components of voter utility such that the more proximity-lead is voting, the larger is β. Finally, δ_j is a vector indicating the effect of background variables z_i on voter utility for party j.

Interpreting β is key, and while voter behavior is represented as a combination of two motivations with β as a mixing parameter, examining the two extremes is helpful for interpretation of the entire range. If voting for a party whose positions differ from those of the voter bears no psychological/representational costs such that voting over policy outcomes is the sole consideration in mind as far as issues are concerned, β will be arbitrarily close to 0, and so the vote will be almost entirely compensational. When the proximity component is dominant β will approach 1, and depending on the institutional context, two interpretations are possible. An empirically large β can be an indication of a voter expressing his or her views and thus voting for the party closest to him or her (representational vote), ignoring considerations of power-sharing and hence the postelection policy-formation process. Alternatively, in majoritarian systems with little bargaining in the policy-formation process, a large β can result from policy-oriented vote (as only little compensation is needed).

To illustrate the calculation in Equation (4), imagine a three-party legislature with parties A, B, and C (s_A=1 − s_B − s_C). In this case,

, and by substituting this into Equation (4), voter i's utility for party A can be expressed as

Calculating the utility for party B as in Equation (5) and taking the difference between the two gives the net utility of voting for A versus B:

Differentiating Equation (5) with respect to p_A and setting the result to 0, we get the optimal placement of p_A for voter i (second-order conditions established in Appendix I):

For a clearer intuition of the solution, consider the two extremes. When β→1 (representational voting), not surprisingly, the prediction of the model reduces to the proximity prediction:

When β→0 (compensational voting) it reduces to

That is, when voting is purely compensational, the ideal placement of party A is the mirror image of policy outcome produced by the combination of parties B and C alone weighted by the impact of party A. Other things equal, the less powerful party A is (the smaller the denominator in Equation [7b]), the farther away it has to locate in order to shift policy outcome in its direction. In addition, holding party impact constant, the more extreme parties B and C are, the more extreme the voter would like party A to locate in the opposite direction with respect to his or her views in order to balance the other two parties. In the extreme case of a β arbitrarily close to 0, the ideal placement of party A is the point in the policy space that yields a policy compromise P identical to the voter's bliss point. Given

(Equation [7b]), that compromise is expressed by

When β is marginally close to 0 (i.e., voting is almost entirely compensational), voter utility still peaks at a point different from his or her own views and then declines. When a party is too extreme or pulls policy “too much” the benefit for the voter declines. Therefore, a voter's decision to vote for a party that is not necessarily ideologically most similar to him or her is moderated by an endogenous feature of my model.⁶

Unlike in the Directional Model, where the moderation depends on the exogenously posed constraint, the region of acceptability, moderation is an endogenous feature of the Compensational Vote Model.

Even a purely compensational voter does not employ a “the more extreme the better” logic.

The mixing parameter represents a summary of a nuanced political reality. In both the theoretical model above and the empirical model below, individuals vary in the extent to which they vote out of representational or compensational considerations insofar as the institutional context in which they vote varies. Under this specification β captures the extent to which voter choice in a given system is motivated by compensational or representational considerations; the model does not allow for heterogeneity among voters in the strategy they employ. Nonetheless, in addition to variation across institutional environments, voter characteristics within an institutional environment may affect the strategies employed. β can be interpreted, then, as an average of voter strategies in a given system.

In addition to a summary of multiple voters, the mixing parameter can be thought of as a summary of multiple selves. As discussed above, each individual carries conflicting considerations; while advancing beneficial policy outcomes, compensational voting for a party whose position differs from the voter's position can be psychologically costly, and conversely, while carrying psychological representational benefits, proximity voting does not necessarily advance one's policy interests. The two motivations often translate to conflicting strategies. The mixing parameter, then (even when estimated at the individual level), also represents the relative extent to which an individual's choice is guided by compensational versus representational considerations.

FROM INSTITUTIONS TO VOTERS: OBSERVABLE IMPLICATIONS

The core of the model leads to my empirical predictions. If, in addition to party platforms, voters are concerned with policy outcomes, they will not necessarily vote for the party whose positions are most similar to their own positions, but rather, they will compensate for postelectoral bargaining resulting in watering-down of their vote and will often prefer parties whose positions differ from their own. Indeed, as I discuss above, because opposition parties affect policy indirectly by placing issues on the agenda, even in a hypothetical case of pure majoritarian regime I expect to find voters employing a mixed decision rule.

How does the argument depend on institutional environments? In my analysis, voters vote over parties available on the ballot but are concerned with policy outcomes. The conversion mechanism from parties on the ballot to policy outcomes varies greatly by institutional context. In some institutional environments political bargaining is an everyday matter, while in others the winner can implement his or her ideal policy with little compromise. Figure 1 summarizes the institutional component of my argument. When institutionalized bargaining takes place voters compensate for the watering-down of their vote by voting for a party whose positions are different from, and are often more extreme than their own ideal points. Since voters utilize their vote to affect policy outcomes, the more power-sharing allowed by the institutional environment, the more voters will compensate for the watering-down of their vote by voting for parties whose positions differ from their own positions. Therefore, empirically, I expect β to decrease with institutional power- sharing.

From Institutions to Voters

This prediction also observationally distinguishes my framework of outcome-oriented behavior from existing theories of issue voting. If voters are concerned with either proximity or direction of platforms, the same regularities should hold irrespective of post-electoral bargaining. If, on the other hand, voters vote to achieve preferred policy outcomes, their taste for compensational versus representational strategies will vary with institutional context as specified above.

Compensational voting is observationally distinguishable from directional voting in three additional ways. The model predicts that the extent to which voters employ compensational strategy depends on institutional context. However, it is likely that the taste for a particular voting strategy also varies across individuals within a given system. Indeed, allowing for individual-level heterogeneity in β, I demonstrate that in the Netherlands, the higher the level of education and the weaker the attachment to a party, the more compensational is one's vote. Behavior of those less educated and strongly attached to a party is more likely to follow principles of proximity voting (Kedar 2003).⁷

In addition, under the Compensational Vote Model, since voter utility for each party—and hence voter choice—depends on predicted outcomes, it depends on the configuration of other parties in a way that it does not under directional voting. As discussed above, a Left-leaning voter might prefer to endorse the extreme left when predicting an outcome to his or her right but might endorse the moderate left (or even a party to his or her right) when predicting that “overshooting” is not necessary (or is even necessary in the opposite direction). Finally, compensational voting does not specify a neutral point that (in one dimension) divides the ideological continuum into two, and therefore, unlike under directional voting, centrist voters will not necessarily be indifferent among parties, in particular, between moderate and more extreme parties.

EMPIRICAL ANALYSIS

The Statistical Model

I derive a statistical model that corresponds with the theoretical model. First, I derive a likelihood function for multinomial choice:

where the dependent variable is vote choice, such that y_ij=1 if the ith voter votes for party j (j=1, 2, …, m), and 0 otherwise, and π_ij is the probability of individual i (i=1, 2, …, n) voting for party j. This probability is a function of his or her utility for that party and his or her utility for all other parties, such that for each voter

where I employ a logistic error structure such that f(a)=exp(a).

Second, the systematic component of the statistical model is in parallel with the theoretical model as it appears in Equation (4); it contains representational and compensational motivations weighted by β and 1−β, respectively, as well as the salience parameter (θ), and m−1 vectors of party-specific effects (δ_j) of the background variables, z_i (δ_j for j=1 is set to 0 for identification purposes), such that each δ_j is a vector including effect coefficients for each variable in z_i, as well as a party-specific constant.¹⁵

The estimated vectors of background-variable coefficients (including party-specific intercepts) vary across polities depending on the relevant cleavages in each particular political system as established in previous research. The models are presented in Appendix II.

This, in combination with Equations (8) and (9), produces a Conditional Logit Model:

Finally, I employ both quadratic and city-block utilities, different model specifications, and different measures of the dependent variable.

The optimization is unconstrained, so in order to compare the relative effects of proximity vs. compensational voting across countries I parameterize β using logistic transformation.¹⁶

In the theoretical model, I defined β₁+β₂=1, so the two parameters are reduced to one, β∈(0, 1). In parallel, in the empirical model I reparameterize β using a logistic transformation such that β=(1+exp(−α))⁻¹. This parameterization ensures that β will be bounded between 0 and 1.

The normalization of β allows me to evaluate whether the data support the theory or disprove it. Estimates of β that are close to the upper bound (high β's with large standard errors) or estimates of β in systems with a high level of power-sharing that are as large as or larger than β in systems with little power-sharing will lead me to infer that the data do not support the institutional hypothesis. On the other hand, estimates of β that are systematically smaller in power-sharing systems than in majoritarian systems will lead me to infer that the data support the theory.¹⁷

To estimate Equations (8) through (10) I multiply through and rearrange terms. I rewrite the bracketed term as

I then multiply through and estimate the model in the form

, which allows separate identification of β and θ. In combination with the description in footnote 16, I maximize the likelihood function with respect to α, θ, and the vector δ.

RESULTS: INSTITUTIONALLY DEPENDENT VOTERS

To test my theory, I first estimate Equations (8) through (10) in each of the four polities. While the issue component of the model is identical across the four, the background variables vary across systems depending on the relevant political cleavages established in previous research. Model specification and comprehensive results of the estimations are presented in Appendix II.¹⁸

.

Table 1 focuses on the estimated β in each of the four polities. Each column presents the results for one of the four polities, and each row indicates the measure used. How do the results vary by electoral system? In almost all cases, vote in the consensual systems is more compensational than vote in the majoritarian systems. Voting in Britain, for example, follows the proximity model more closely than voting in the Netherlands across all measures (0.833, 0.730, and 0.849 in the former and 0.543, 0.543, and 0.597 in the latter), and the same pattern evolves from a comparison of the Netherlands with Canada (0.770, 0.613, and 0.880 in the latter). The results for Norway are somewhat weaker yet still in the expected direction. The results produced by utilizing the seat-share measure indicate that voting behavior in Norway is more compensational than in Britain, and, similarly, results produced by the mixed measure indicate that voting in Norway is more compensational than in Canada. The rest of the results for Norway do not allow me to generally infer that voter behavior in Norway is different from that in Canada and Britain (the confidence intervals overlap substantially). Overall, however, with one exception (vote-share measure in Norway and Canada), all confidence intervals in the power-sharing systems are centered around smaller βs than in the majoritarian cases. In other words, the pushing and pulling, bargaining, and compromise—all consequences of institutional features—are reflected in voter behavior.¹⁹

In addition, I conduct a set of robustness tests for the empirical predictions of the model. These tests include variations on the specification reported above. In particular, I repeat the test for each of the three measures of party impact varying: (a) the utility function from quadratic to city block, (b) the dependent variable to thermometer rating for parties, (c) the measure of party position to average perceived position, and (d) the addition of retrospective voting and candidate evaluation as further covariates. I also conduct the tests with combinations of these changes: (e) city block—thermometer ratings, and (f) thermometer rating—average perceived party placement. Overall, the original results and these tests provide 21 different estimates of β for each polity (with the exception of the Netherlands, where they provide 14 such estimates).

In all specifications except c the point estimate of β in either of the majoritarian cases is greater than the estimate in either of the proportional cases on at least two of the measures. In the cases of city-block utilities, thermometer ratings for parties, and the combination of the two, the results hold for all three measures, and in the cases of the original model, additional covariates, and the combination of party thermometers and average perceived party position, they hold for two of the three measures and for three of the four polities in the third measure. All these results are in the expected direction, but while most of these differences are statistically significant, in some cases they statistically overlap.

The results of specification c show no systematic pattern. Examination of these results in combination with specification f (where the results are in the right direction but are not statistically significant for two of the three measures) suggests that the measure of party placement as average perceived placement by voters is the reason for the weak results. As discussed above, and as Blais et al. (2001), and Westholm (1997) argue, given that the focus of this study is individual voter's choice, the appropriate measure of party position is the position as it is perceived by the individual voter, rather than an average.

To complement these figures, Figure 2 presents density plots of the four estimates of β based on the seat-share measure.²⁰

Issue Voting in Four Systems

In comparing the βs across the models I assume that the estimates of β in the different countries are independent of one another conditioning on the covariates in each model. This implies, for example, that the taste of Dutch voters for compensational/representational strategy in the Netherlands is independent of the taste of Canadian voters for these strategies. Indeed, it is hard to imagine why there would be dependence in taste for vote-choice strategy among voters in different polities, conditioning on systematic factors of voter choice. Relying on this assumption, and in addition to the analyses above, I compute the differences in β between each majoritarian case and each proportional case (not reported). The results are consistent with the results presented in Table 1. All the point estimates of the differences are positive, indicating that β in the majoritarian systems is greater than β in the proportional systems. The Netherlands is consistently distinguishable from both Britain and Canada. Norway is not always statistically distinguishable from the majoritarian systems, although the direction of the results is as expected.

Does the Proximity Model sufficiently capture the information in the data? To further address this issue, I compare the results above to a restricted model consisting of the proximity component and the background covariates.²¹

That β is removed from the upper bound not only in the power-sharing systems but also in the majoritarian systems is particularly interesting. Although Britain and Canada are extreme cases of the majoritarian models, they are not ideal types of the model (see Lijphart 1999, 10–21, for detailed discussion of Britain, and Figure 14.1 there for placement of the four cases on the Executive-Parties dimension). Even in First-Past-the-Post systems, where opposition parties do not participate directly in the policy-formation process, they indirectly affect policy outcomes by shaping the political discourse and forcing other parties to adapt their policies (Avakumovic 1978; Meguid 2002). By the same token, the results presented in Table 1 show that β in the proportional cases is safely away from the lower bound—issue voting is not entirely compensational even in the proportional cases. This is consistent with both the theoretical expectation and the nature of the empirical cases; similar to Canada and Britain on the majoritarian end, both Norway and the Netherlands are strong, yet not pure, cases of consensual democracy. Finally, a comparison of the cases from the 1980s (Britain and Norway) with those from the 1990s (Canada and the Netherlands) does not reveal a systematic pattern with respect to timing.

These results shed light on the institutional component of the argument and on the extent to which the Proximity Model captures voter behavior. Voters in the power-sharing systems behave differently from voters in the majoritarian systems. Under the former, voters compensate for the watering-down of their vote, often voting instrumentally for a party whose positions differ from their own. Under the latter, votes are hardly diluted, and voters make their choice accordingly.

DISCUSSION AND EXTENSIONS

How do voters' positions on issues translate into vote choice? The common wisdom of issue voting holds that voters make their choice based on their evaluations of party positions (platforms). Since, according to this framework, platforms are the object of voter evaluation, any given theory of voter choice based on platform evaluation should account for voter behavior irrespective of the way votes are converted to policy.

I demonstrated that this is not the case. Regularities of voter behavior vary by institutional context. In particular, voter choice reflects variation in postelection bargaining. This finding suggests that something else is at work, namely, outcome-oriented behavior. My framework is not only consistent with the unaccounted-for variation observed in current studies, it also allows me to reinterpret (rather than declare a winner in) the current debate between the two leading theories of issue voting, proximity and directional voting.

In itself, the notion that voters are concerned with policy outcomes is not a novel insight (Austen-Smith and Banks 1988; Baron and Diermeier 2001). However, my framework allows me to explain how variation in institutional mechanisms, types of democracy, and even individual-level characteristics account for variation in strategies voters employ. This article has focused on the first component; I have shown that outcome-oriented motivation leads voters in parliamentary regimes to endorse parties whose positions differ from the voters' own positions depending on the institutional environment in which voters make their choice.

As formalized in this article, the theory is captured by a single parameter. But as mentioned above, β stands for a complex reality. One can estimate, for example, the extent to which each voter's motivation is compensational or representational, the extent to which voters are compensational or representational on different policy dimensions, the extent to which supporters of different parties are motivated by compensational or representational considerations, and the like.

In addition, the argument travels across types of democracy. In fact, it applies to any political system where no single player can implement his or her preferred policy in its pure form. While the bargaining units in parliamentary systems are parties (or subsets of members of the parties) in the parliament, in presidential systems the power is distributed between the executive and the legislature, and in federal systems bargaining takes place between regional and federal agents.

Indeed, scholars of presidential systems have examined policy balancing by voters in both general and midterm elections (e.g., Alesina and Rosenthal 1995 and Mebane 2000 for analysis of the American case). As some of these studies argue, depending on the voter's position, the more extreme and/or more powerful the president is, the more extreme and/or powerful the voter prefers Congress to be in order to balance the president. But policy balancing in the United States is a specific case of policy balancing by voters in general—voters incorporate postelectoral bargaining into their choice: their vote in legislative elections depends on their prediction of (or in Midterm elections, knowledge of) partisan control of the executive.

While in presidential systems voters balance the legislature vis-à-vis the executive, in federal systems the branches balanced are regional and federal governments. Examining electoral returns from almost three decades in Germany (Lohman, Brady, and Rivers 1997), and adding analysis from the postunification era (Gaines and Crombez 2004, Kedar 2004), scholars have demonstrated that the party in control of the chancellery systematically loses seats in subsequent state elections. While partisanship and the economy explain part of the gap, controlling for differential in turnout between elections at the two levels, these studies demonstrate that voters use Land elections to counterbalance the federal government and, thus, engage in vertical policy balancing.

Whether parliamentary, federal, or presidential forms of power-sharing, the conceptual difference between the analysis of voter choice in this study and the classic formulations of voter choice has to do with the transformation from the set of possible parties (platforms) to the set of possible policy outcomes. This study shows that while voters choose a party (a platform) from a set of competing parties, they are often concerned with policy outcomes. The complexity of policy formation and the mapping from vote choice to policies often result in voter preference for a party that is not necessarily the most ideologically proximate to the voter. Unpacking the fashion in which institutional mechanisms convert votes to policy outcomes is a crucial step toward better understanding of voter choice.

APPENDIX I: ANALYTIC SOLUTION

Second-order condition on

with respect to p_A:

Rewriting Equation (A1), it is easy to see that this is a maximum:

Since 0<β<1 and θ>0, both expressions are negative, and therefore

. Differentiating

with respect to p_B, we get

Setting the result to 0 and solving for p_B:

Second-order condition with respect to p_B:

Rewriting Equation (A4), it is easy to see that this is a minimum:

Since all elements are positive and 0<s_B+s_C<1,

. These results are symmetric with respect to p_C.

APPENDIX II: COMPREHENSIVE RESULTS

Table A1 presents results for the power-sharing cases—the Netherlands and Norway—and Table A2 presents results for the majoritarian cases—Britain and Canada. Each model introduces an estimate for compensational–representational voting, salience, and party-specific effects of the background covariates (as in Equation [10]). Coefficients of the covariates represent effects of the respective variables on the log odds of voting for the relevant party compared to the reference party (coefficients of the reference party are normalized to 0).

Model specification for each polity depends on the relevant political cleavages established in previous research. Model specification for Canada is based on Clarke, Kornberg, and Wearing (2000) and Johnston et al. (1992). The results include all regions but Quebec, which I estimate separately. Results for Quebec (not reported here) are similar to the general results presented here. Model specification for the Netherlands is based on Dorussen and Taylor (2001), Quinn, Martin, and Whitford (1999), and Taylor and Dorussen (2000). The model for Britain is based on Alvarez, Nagler, and Bowler (2000). I also draw on Whitten and Palmer (1996) for model specification in the British and the Dutch cases.