Voting polarization on redistribution across democracies. John D. Huber Piero Stanig

Transcription

1 Voting polarization on redistribution across democracies John D. Huber Piero Stanig Department of Political Science Columbia University New York, NY Abstract. We explore how the social and political context affect voting polarization by income, which is the difference in the propensity of poor and non-poor voters to support right-wing parties that espouse low taxes and low levels of government services. To this end, we draw on a range or arguments about the dimensionality of the policy space, the salience of redistribution issues, and the impact of the party systems on voting cross-pressures to develop expectations about the types of countries that should have the highest and lowest levels of voting polarization. We test these arguments using survey data from the Comparative Study of Electoral Systems. Prepared for the 2006 Annual Meetings of the American Political Science Association, Philadelphia, PA.

2 Introduction A central premise underlying models of democracy is that the issue of income redistribution drives electoral competition. A polity consists of the haves and have nots, or the rich and poor. The poor wish to use elections to support parties that will tax the rich and give the money to the poor. The rich wish to avoid taxes and redistribution. This simple and powerful tax and transfer framework for thinking about democracy has allowed scholars to explore a wide range of important issues. In their classic work, Meltzer and Richard (1981) use the tax and transfer framework to study the relationship between income inequality and taxes. Scholars have built on this framework to study an enormous array of other questions, such as transitions between democracy and authoritarianism (Acemoglu and Robinson 2006, Boix 2003), the effect of income inequality on economic growth (Alesina and Rodrik 1994 ), the impact of skill specificity on preferences for social protections (Iversen 2005), the impact of religion (Scheve and Stasavage 2006, Benabou and Tirole 2006) and social mobility (Benabou and Ok 2001) on preferences for redistribution, the impact of individual mobility and interjurisdictional competition on redistribution (Epple and Romer 1991), the impact of electoral laws and separation of powers on economic performance (Persson and Tabellini 1999; 2000), and the impact of electoral laws on left-party dominance and redistribution (Iversen and Soskice 2006). Yet sitting somewhat uncomfortably along side this array of political economy research based on the tax and transfer framework is another body of research that forces one to ask whether the assumptions on which this framework is based make sense. For a number of years, for example, sociologists have described the decline in class voting across advanced democracies. Whereas thirty years ago, working class individuals voted left and others voted right, this no longer seems to be the case. Today, class is a relatively weak predictor of the vote (Hechter 1

3 2004). 1 In American politics, there is a great deal of interest in understanding to what extent poor voters support the Republican party because of concerns about issues other than redistribution (CITES). And for many years, Inglehart has argued that material issues (like redistribution) have declined in importance to voters, and have been replaced by post material concerns such as the environment and participation. We ve also seen a recent renewal of interest in religion, with scholars seeking to understand how and whether religious attitudes encourage individuals particularly poor ones -- to take political positions that run counter to their tax and transfer interests (e.g., Scheve and Stasavage 2006, forthcoming; De La O and Rodden 2006) The goal of this paper is to understand the covariates of the differences in the propensity of poor versus non-poor voters to support right wing parties. Specifically, we explore how the social and political context affect what we call voting polarization by income, i.e. the difference, all else equal, across income groups, in the propensity to vote for right-wing parties that prefer low taxes and low levels of government services. To this end, we draw on existing research that seems relevant to understanding why poor voters might resemble middle class and rich voters in their support for right wing parties, and why on the other hand they might behave differently. We use this research to develop a somewhat extensive number of arguments about which types of political systems, and which types of social contexts, should lead to more or less support by poor voters for right-wing parties and as a consequence to higher or lower degrees of polarization between income groups. We test these arguments using the Comparative Studies of Electoral Systems data. The paper is organized as follows. In the next section, we provide some definitions that are required to develop our argument. In section two we analyze how income and religious participation affect both attitudes regarding redistribution and voting behavior. In section three we survey the existing literature and derive some testable predictions. We describe the variables 1 For critical reviews regarding the decline in class voting, see Evans (2000) and Manza, Hout and Brooks 2

4 and the estimation technique in section 4, and present the results in section 5. We draw some tentative conclusions in section 6. Some essential definitions To study the impact of being poor versus non-poor on the support of right-wing parties requires that we define poor, non-poor, and right wing. In the CSES data, income is reported in quintiles, and we define poor as the bottom quintile, rich as the top quintile, and middle income as the three middle quintiles. We impute missing income data. Our specific goal is to understand the different propensity of voters in each income group to vote for parties that favor lower taxes and less redistribution. It is important to recognize that voting right in this narrow tax and transfer sense can be quite different than voting right on some general left-right dimension. The problem with using a general left-right variables is that the substantive meaning of left-right varies a great deal across countries, and often has little to do with economic redistribution. We use the data from Laver and Benoit (2006) to place the parties that individuals vote for in the CSES national election studies on a 20-point economic left-right scale. Dimension 1 from this data set is the closest measure we have found to positioning parties on the tax and transfer scale. For this variable, country experts place parties on a scale ranging from 1 (party Promotes raising taxes to increase public services ) to 20 (party Promotes cutting public services to cut taxes ). We code a voter as having voted for a rightwing party if he or she casts a vote for any party with a score of 10.5 or greater on this 20 point scale. 2 Using the Benoit and Laver data, it is easy to show that the general left-right scale, while clearly related to the redistribution scale, is not the same thing. In Figure 1, we plot the party positions on the redistribution dimension against their placement on the general left-right (1995). 2 We also use a different cut-point 12.5 to assess the robustness of our statistical results. 3

5 dimension, according to the Laver and Benoit (2006) dataset. As one can observe, there is quite a good deal of variation in how close the parties are to the 45-degree line, which marks perfect congruence of the two scales. And there are a large number of parties that are in the wrong quadrant, i.e. they are left of center on a general dimension but right of center on the economic dimension or vice versa. This is especially true of parties that are right-wing on the general leftright scale many of them are left-wing on the redistribution scale. It is clear, then, that it is better to use a specific redistribution scale than a general left-right scale if one wishes to understand the propensity of poor voters to support parties that are against taxes and redistribution. Attitudes toward redistribution and right wing voting by rich and poor One reason why poor people may not differ much from non-poor voters in their support of economic right-wing parties might simply be that the attitudes of the poor are more right wing on taxes and transfers than we commonly assume. While it is impossible to use the CSES data to examine this question, before examining at actually voting, it is useful to consider briefly the actual attitudes of poor voters toward the core issues of redistribution and the government s role in the economy. We consider two issues from the World Values Survey that were asked across a wide range of countries, and that tap attitudes that should be related to the core tax and transfer assumptions. The first concerns inequality, and the second individual responsibility. The inequality survey question reads as follows: Now I'd like you to tell me your views on various issues. How would you place your views on this scale? 1 means you agree completely with the statement on the left (Incomes should be made more equal); 10 means you agree completely with the statement on the right (We need larger income differences as incentives). We code responses of 7 or higher as being right-wing and responses of 4 or lower as left-wing (with responses of 5-6 coded as neither left nor right). 4

6 The individual responsibility question reads as follows: How would you place your views on this scale? 1 means you agree completely with the statement on the left (People should take more responsibility to provide for themselves).; 10 means you agree completely with the statement on the right (The government should take more responsibility). We code responses of 1-4 as right wing, and responses of 7-10 as left-wing, with responses of 5-6 coded as neither right nor left. Figure 2a shows the proportion of poor, middle income, and rich respondents who take right-wing positions on the inequality question. 3 The countries are sorted according to the proportion of poor respondents who are right wing. Several points emerge from the figure. First, in most countries, there are substantial proportions of poor people who take the right wing position on inequality. In only three countries is the proportion of poor voters who take rightwing positions on inequality less than one-quarter, and in nine of the 24 countries, more than 40 percent of poor people adopt right wing positions on inequality (topping out at 55 percent in Korea). Second, rich respondent are almost always more right-wing than the poor on the question of inequality. Only in Romania are the poor more right-wing than the rich. Third, the degree to which rich and poor are polarized in their attitudes on inequality varies a great deal. In some countries (e.g., Chile, Belgium, Taiwan and Finland) the difference between the proportions of rich and poor respondents taking right-wing positions is very small (less than 8 percent), whereas in other countries (e.g., Netherlands, Ireland, Mexico, Czech Republic, New Zealand, the UK and Bulgaria) the differences are substantial (greater than 20 percent). Figure 2b examines left-wing attitudes on inequality. The data are sorted according to the proportion of rich who are left-wing. Just as the poor frequently cross-over and take the rightwing position on inequality, so do the rich cross-over and take left-wing positions. But they do so in much smaller proportions, with only three countries having more than 40 percent of the rich 3 To ensure comparability in the set of countries used to study attitudes and voting, the data WVS countries displayed are limited to those countries that also exist in the CSES. 5

7 taking the left-wing position. As in the examination of the right-wing positions, we find the relative proportions of rich and poor taking the left-wing position to be consistent with the core PE assumptions: in every single country, the proportion of rich who are left-wing is less than the proportion of middle-income or poor who are left-wing. And we also find that the polarization between rich and poor attitudes varies dramatically across countries, with the difference being greater than 20 percent in Poland, Mexico, the UK, Switzerland, the Czech Republic, Bulgaria and Slovenia (where polarization is at its maximum, at 29 percent), and the difference being less than 8 percent in Romania and Taiwan. In 8 of 24 countries (Korea Taiwan Netherlands, the United States Ireland Japan Iceland and Poland), the proportion of poor who take right-wing attitudes on inequality is greater than the proportion of poor who take left-wing attitudes. Among the rich, however, in only four countries (Slovenia, Finland, Chile and Romania) is the proportion with left-wing attitudes greater than the proportion with right-wing attitudes. Figures 3a and 3b present analogous graphs of attitudes toward individual responsibility. As in the question regarding inequality, in most countries, there are substantial proportions of poor individuals who hold right-wing attitudes toward individual responsibility, yet there are substantial differences across countries in the proportions who take right wing positions, and in the polarization between rich and poor on this issue. We also find that the rich usually take the right-wing position in higher proportions than the poor, though this is not true in five countries. It seems clear, then, that among rich and poor, there are many individuals who do not have the attitudes toward inequality and individual responsibility that one might expect. Based on this fact alone, we might expect that there are substantial numbers of poor people who support right wing parties, and substantial number of rich people who support left-wing parties. Indeed, this is the case. Figure 4 shows the proportion of poor, middle income and rich respondents that voted for a right wing party, in the countries that we included in the regression 6

8 analysis we present below. The countries are sorted according to the proportion of poor voters that voted for a right wing party, which ranges between the low extremes found in some former Communist-ruled countries (Hungary, Romania, and Poland) to more than 40 percent in Australia, Iceland, Bulgaria and Ireland. In the majority of countries, the respondents in the lowest categories of income vote for the right less often than people in higher income categories: voting behavior shows a consistent pattern according to which a smaller proportion of poor respondents than of rich and middle class respondents supports parties that propose right wing platforms on the issue of income redistribution. This pattern is similar to the one found in the case of attitudes regarding inequality. The exceptions to this pattern are Romania, where poor voters are more likely to have voted for a right wing party than rich voters, but less likely than middle class voters; France, where poor voters are slightly more likely than the middle class, but less likely than rich voters, to support the right; Portugal, where poor voters are much more likely than middle class individuals and much less likely than rich voters to support a right wing party; and Ireland, where poor voters are the most conservative of the three income groups. The degree of polarization varies a lot across countries: in some countries the proportion of rich and poor voters that support the right differ by around twenty percent (United States, Sweden, Norway); in other countries the difference in the proportions is negligible, less than five percent (France and Switzerland, Hungary in 1998, Switzerland in 2003). Interestingly, the voting behavior patterns follow to some extent the pattern related to attitudes, but the mapping is not that close. Indeed, it is true that the country in which poor voters are the most conservative group when it comes to voting, Ireland, is one of the 8 (out of the 24 analyzed above) countries in which more poor people have right-wing than left-wing views on inequality; yet, it is also one of the most polarized countries when it comes to attitudes. Contrary to the conventional wisdom in the US that many poor people vote right, the US is the country 7

9 where income polarization in voting is greatest, but is also one of the countries in which the poor are most conservative regarding redistribution. Clearly, the cross- country variation in voting behavior, in the sense of voting for parties that propose economically conservative platforms, is not explained in a trivial way by the differences in attitudes on inequality and responsibility. In the next section, we sketch arguments that might help us to understand how context affects differences between rich and poor voting. Where should one expect voting polarization by income? We wish to explain cross-national differences in the relation between income and voting behavior. Why do poor people who would benefit from redistribution vote for parties that do not support it? We are specifically interested in voting polarization by income. Such polarization increases when the probability of the poor supporting the right diverges from the probability that other income groups will support the right. In this section, we draw on a variety of existing arguments in the literature to develop testable hypotheses about factors that affect voting polarization by income. The arguments are all related to factors affecting the relative salience of the redistribution dimension in a given society. We wish to understand factors that are associated with a salient second dimension that can affect vote choice by the poor, and factors that affect the salience of the redistribution issue, itself. In addition, we want to understand the extent to which the party system forces voters to make tradeoffs across dimensions. The presence of a second dimension. One of the central arguments about why poor voters might support parties that are right-wing on economic issues focuses on the fact that the policy space is often multidimensional. If the preferences of poor voters on economic policy are less important than preferences on some other dimension such as race or religion then the 8

10 poor voters might abandon the party that supports their economic preferences for a party that supports their preferences on the second dimension. This will only occur, of course, if the party system has no parties that espouse the platform preferred by poor voters on both salient dimensions. This second dimension argument has been analyzed by many scholars, and developed formally by Roemer (1998, see also 2001). Voting polarization by income should therefore be highest when the party system does not force voters to choose between different parties on different dimensions. Voting polarization by income across countries should therefore be related to factors that affect whether there actually exists more than one salient dimension. There is a long tradition in comparative politics of focusing on religion as a salient second dimension. Lipset and Rokkan s (1967) classic study was perhaps the first to recognize that in many countries, an enduring religious cleavage emerged between Catholics and Protestants, between religious and nonreligious individuals, or both. These cleavages led to alignments between religious individuals and religious parties. Since religious parties are almost invariably right-wing on the taxes and redistribution dimension, the economic cost of religiosity is born by the poor, who must crossover and support right-wing parties if they wish to vote based on religion. Thus, there should be less polarization by income in the most religious societies, as more poor will crossover to the right. This holds only as long as poor people exist who vote based on their religious preferences dimension rather than their redistribution preferences. Religiosity might also affect voting polarization by influencing whether poor religious people feel they actually face a trade-off between economic and religious dimensions. Some scholars argue that being religious directly influences ones attitudes toward economic redistribution. Specifically, scholars have argued that religious individuals often place a high value on individual responsibility, and this leads them to have right-wing preferences on the dimension of income redistribution (Guiso, Sapienza, and Zingales 2002). Scheve and Stasavage 9

11 (2006) argue that this is true because religion provides a psychic benefit that ensures religious individuals against the risk of personal economic distress. Benabou and Tirole (2006) suggest that this might happen because religious beliefs in an afterlife reward, correlated, in a Weberian fashion, to the effort one exerts during one's life, induce religious people to work harder. Therefore religious individuals do not want the fruits of hard work to be redistributed to less hard-working non-religious people. At the same time, a low tax rate induce people to cultivate their religiosity and to provide religious education to their offspring: indeed, in a low redistribution political economy, effort is more valuable, and religious beliefs increase an individual s willingness to exert effort. In the model by Benabou and Tirole, two equilibria exist: one in which people have a "Protestant work ethic", exert high effort and prefer low taxation and redistribution; one in which agnostics, with preferences for a high degree of redistribution, are prevalent. There are a number of empirical and theoretical arguments, then, asserting that individuals who are religious are more likely to be conservative on economic issues than individuals who are not religious, which implies that voting polarization by income should decline as the religiosity of a society increases. This occurs not, from this religion and attitudes perspective, because individuals face a tradeoff across two dimensions, but because religiosity actually decreases the dimensionality of the policy space by aligning the redistribution and religious dimensions. The power of the church in society should affect the extent to which religion is a salient second dimension for poor voters, even when these poor voters are not religious. In some societies, the church is a powerful social actor that creates access to jobs, provides important social benefits, and/or provides important social networks. Where the church plays a strong social role, many poor individuals will have incentives to engage actively in church participation, independently of whether they believe or not (Huber 2005). Since strong churches are typically associated with parties on the right, we might expect poor individuals to feel most cross-pressured 10

12 to vote for parties on the right in societies where churches are strong. They may do so, even if they do not accept church religious doctrine, through a captive audience or propaganda effect (their participation in worship services and church networks socializes them to support right wing parties) or because they believe that voting for a right-wing religious party will contribute to the vitality of the church, thereby helping to ensure future access to the non-spiritual benefits that churches provide. By this logic, the more powerful the church in society, the less voting polarization by income there should be. The structure of religious institutions should affect voting behavior in a second way: in countries that are highly homogenous from the point of view of religious denominations, the church organization might be a powerful lobby that transmits a consistent message and makes the religious dimension more salient. In a country where a range of religious denominations exist, the ability of the churches to lobby and be politically active is reduced because they represent a plurality of voices. We then expect that, all else equal, in societies with a more homogenous religious landscape, the second ( values ) dimension should be more salient than in countries in which churches speak different voices. Right-wing populism can also be a sign of a salient second dimension. Extreme right parties make nationalist appeals that stress anti-immigration and anti-multiculturalism policies while also using an anti-establishment and anti-system rhetoric. Such a rhetoric appeals to poor voters, who provide a strong basis of support for such parties (e.g., Hainsworth 2000, Ignazi 2004). Although the support for such parties is not large in most countries, it is reasonable to expect that the presence of such parties indicates the salience of a nationalism dimension, giving poor voters an incentive to vote for the right. We should therefore expect that the presence of extreme right-wing parties with enough support to survive should be associated with decreased voting polarization by income. 11

13 By a similar logic, ethnic divisions within societies should also force poor voters to choose between economic redistribution and a second salient dimension in politics. When societies are divided by race or language, poor voters may be forced to choose between supporting a mainstream left-party and a right-wing party that makes ethnic appeals. This would imply that ethnic divisions would lead to less voting polarization by income. The importance of redistribution. The tradeoff voters may face between the economic dimension and other dimensions will also be less important if the salience of the economic dimension declines. Inglehart (1990) argues that as societies develop economically, traditional concerns with material needs recede and are replaced by new values issues, such as individual participation, women s rights, and the environment. Typically, parties of the left lead the way in addressing these issues, causing many middle class voters to cross-over and vote for the left. Other scholars argue that economic modernization produces cognitive mobilization, which leads to a decline in class voting. To the extent that class is correlated with income, a decline in class voting would lead to lower voting polarization by income. Finally, others have argued that as societies become more modern and the size of the welfare state increases, the number of poor voters declines. This leads parties to focus on issues of concern to the middle class at expense of issues of concern to the poor. For all these reasons, we might expect that as economic modernization increases, voting polarization by income should decline because the relatively poor will be less concerned with economic redistribution, and the non-poor will be more inclined to support parties on the left. Levels of urbanization should also influence the values of poor voters. Scholars have shown that urbanization is associated with strong left-wing values, with an emphasis on social protections, class, as well as liberal social attitudes (Rodden 2005). To the extent that a country is more urbanized, poor voters will be concentrated more in urban areas, and will be expected to have left-wing values on economic and social issues, increasing voting polarization by income. 12

14 The salience of the economic dimension will also be affected by the differences between the left and right parties on economic (tax) policy. McCarty et al. s (2006) study of the United States underlines the importance of the ideological polarization of the party system. As America s two main parties diverge from each other, the differences between them in expected tax and transfer policies will increase. This makes it more attractive for poor voters to support the Democrats and non-poor voters to support the Republicans. We should expect to observe more voting polarization by income in countries where the party system is most polarized (i.e., where the distance from the left to right is greatest), and the costs of voting for the wrong party on the economic dimension are higher. Like ideological polarization, a country s history with a Communist Party regime should also affect the risk and reward of voting for the right. Former Communist-run countries began with a very egalitarian distribution of income, and transitions to a market economy often created substantial gaps in income and wealth between the winners and losers of the market transition. In such countries, supporting parties that advocate the most liberal market economy will pose extraordinary risks to the poor relative to supporting parties on the left typically former communists that oppose the market economy, or advocate market economies with substantial safety nets for the potential losers. To the extent that the poor are most vulnerable to the market reforms, we should expect voting polarization by income to be higher where there is a historical legacy of a planned economy and a Communist-run regime. Economic openness should also affect the salience of economic redistribution issues. If a country s economy relies to a great extent on world markets for the sale of its products, the working class is often most vulnerable to the oscillations of the international markets. In such situations, voters in the lower income categories should favor even more redistribution, in the form of safety nets. By contrast, capital in open economies will fear the impact of high taxes and redistribution on international competitiveness, and thus will be opposed to redistribution 13

15 (except where capital is dependent on the investment by labor in specific skills, see Mares 2003). We should therefore expect voting polarization by income to increase as trade becomes more open. Finally, the salience of the income redistribution issue for poor voters may depend on social structure. Some scholars have identified a negative exposure effect, where group based differences lead to a decline in support for income redistribution. The idea is that preferences regarding redistribution have an interpersonal component, and that as the number of recipients in a community go up, support for redistribution goes down in general, but does so particularly when resources are transferred to members of another group (Luttmer 2001; see also Alesina et al. 2001) Unless redistributive policies can be designed in ways in which they only target the ethnic in-groups 4, people should oppose redistribution if they expect that income is going to be redistributed to members of a different ethnic or racial group. Support for redistribution should therefore decline in societies that are most divided on racial or ethnic lines. Party systems and trade-offs across dimensions. We care not only about the salience of two dimensions, but also about whether the political system forces poor voters to choose their party based on one dimension or another. De la O and Rodden (2006) argue that (compared with majoritarian systems) in proportional representation systems, voters are more likely to vote on social issues because they do not face a tradeoff in their vote choice on economic issues (because there are left-wing parties on economics that are left or right-wing on other dimensions). Of course, the same conceptual argument should work in reverse voters can vote on economic issues because they do not face a trade-off on other dimensions (because there are left and rightwing parties on these other dimensions). It is unclear to us, then, why the nature of the party system affects which dimension is most salient to voters. But whether or not such forced choice 4 This might be the case in countries that are ethno-federal, like, among the ones we analyze, Switzerland, Canada and Belgium. We leave this question to future research. 14

16 exists should affect voting polarization by income regardless of whether voters care more about economic or social policies. Political systems with a low number of parties may force voters to choose between parties based on a trade-off across issues. In the US, for example, the Democrats are liberal on economic policy and on policies that affect individual rights and liberties, while the Republicans are conservative on both dimensions. Poor voters must therefore vote right if they wish to limit individual rights in issue areas like abortion or gay marriage, and if they care more about such issues than economic ones. The same dynamic could affect cross-over voting by non-poor, who must vote left if they want to protect individual rights and care more about these individual rights issues. On the other hand, if an electoral arena allows a voter to choose among parties with any combination of positions on the most salient dimensions, the voting polarization on income should increase because poor voters will not need to vote against tax and transfer parties. If the US, for example, had a viable party that was to the left on the economic dimension but to the right on the issues related to individual rights, the question of crossing over would not confront poor voters. As a consequence, if we assume (a) that poor voters prefer left parties on economic issues and that non-poor voters prefer right parties, and (b) that some dimension other than economic redistribution is most salient for some poor and non-poor voters, then there should be large voting polarization by income when the party system allows all types of parties to form. In such situations, voters can always vote according to their preferences on the economic dimension, independent of the salience of other dimensions. Testing the arguments 15

17 To estimate how the macro, country-specific factors affect the polarization in voting behavior across poor and non-poor voters, we follow the two-stage estimation techniques described in Huber, Kernell and Leoni (2005). In the first stage, we estimate a separate probit model in each country. In these models, the dependent variables take the value 1 if the voter voted for a rightwing party. The independent variable of central interest, Poor, is an indicator variable that takes the value 1 if the respondent s income is in the bottom quintile of the income distribution. In our second stage regressions, we use the Poor coefficients as the dependent variable, and attempt to explain cross-national differences in these coefficients as a function of the macro variables. The second stage regressions are weighted using the standard errors on the Poor coefficients from the first stage regression. The weighting strategy we employ is taken from Borjas and Sueyoshi (1994). Leoni (2005) uses Monte Carlo simulations to demonstrate that this two-stage approach to the analysis of survey data with probit has desirable properties when sample size in the first stage is relatively large (as is the case for our country surveys). In addition to the Poor indicator variable, all first-level models also include a number of control variables: age, age 2, education, a female indicator, and a number of indicator variables for occupational status (housewife, student, unemployed, retired, and other employment, with the omitted category being full time ). We also include an indicator variable for the respondent s religiosity. Weekly attendance takes the value 1 if the respondent reports attending religious services at least once a week and zero otherwise. Unfortunately, since the question about religious participation was not asked in all surveys, including this variable in the first stage causes us to lose almost a quarter of the country-observations in the data set. For most of what follows, we have second-level observations, which include 20 different countries. It is possible to include other income variables in the first stage, which changes the interpretation of the Poor coefficient. If the first-stage models include a Rich indicator variable (top quintile in income), the Poor coefficient tells us how much more or less likely a poor person 16

18 is to vote right than a middle income person (quintiles 2-4, which is the omitted category). If we include a Middle income indicator rather than Rich (so Rich is the omitted category), the coefficient on Poor tells us much more or less likely a poor person is to vote right than a rich person. In what follows, we will discuss the robustness of the results to these various possible first-stage specifications. The first stage model that we will be our central focus, however, includes the Poor indicator variable, no other income variables (so we are simply comparing poor to non-poor), the religious attendance variable, and the other controls. It is impractical to report all of the results from the (up to) 38 country-specific regressions), which would entail reporting 418 coefficients and standard errors. But in the baseline regression model (that include the religion variable), the coefficient on Poor ranges from -.74 in the Czech Republic (2002 election) to.29 in Switzerland (1999 election). Figure 5 shows the coefficient estimates of Poor and its confidence interval for each country. The mean coefficient is -.15, the coefficient is negative in almost all countries (25 countries of thirty), implying that in most countries, controlling for other variables, including religiosity, poor are more likely to support left parties than are non-poor. The second-stage, system-level variables. The argument in the preceding section suggested a large number of variables that could be related to the propensity of poor voters to support right-wing parties. A number of these a related to religion. To test the argument that a strong religious cleavage decreases voting polarization by income, Weekly attenders(%) measures the proportion of CSES respondents in the country who report that they attend services at least once a week. In the first-stage regressions where we include religious attendance, it is unlikely that we will find an effect of this variable, although it may the case that there is a network effect in the sense that in countries where a large proportion of people who attend church, there may be a propensity of poor voters to vote right independent of whether they attend church themselves. To measures the power of the church, False witnessers is the percentage of 17

19 individuals in the country who attend religious services weekly, even though they report they do not believe in heaven (i.e., it is the number of respondents who do not believe in heaven and attend weekly, divided by the total number of respondents who do not believe in heaven). The variable takes a higher value when individuals feel it is important to attend services, even if they do not believe, which we argue should be related to church power in society. The data are taken from the World Values Survey (IV). The final religious variable is Religious fractionalization, a standard fractionalization measures that can range from 0 to 1. It is a measure of how likely it is that two randomly chosen individuals belong to different religious groups, and is taken from Alesina et al. (2003). If religious fractionalization works the same way as ethnic fractionalization, then more religious fractionalization should lead to less polarization (because religion would be more likely to pose a second dimension, as religious groups compete with each other, or because of the negative exposure effect: the members of one religious group do not want to pay taxes for another group). We might expect, however, that religious and ethnic fractionalization do not work in the same fashion, if, as noted above, a second dimension is more salient when religious organizations speak in a unitary voice than when they represent a plurality of positions regarding political or politicized issues. Two other variables should reflect the salience of a second dimension. Extreme Right takes the value 1 if there exists a party that is positioned at 17 or higher on the Benoit-Laver immigration scale. The variable should have a positive coefficient if the poor are seduced to support the right by the nationalist appeals of extreme right parties. Ethnic fractionalization measures how likely it is that two individuals are from different ethnic groups, where ethnicity can be either language or race. The data are from Alesina et al (2003). It should have a positive coefficient if the negative exposure effect increases poor support for right-wing parties, or if the number of issue dimensions increases with the number of groups. 18

20 Several others variables are related to the salience of the redistribution issue. GDP/capita (ln) is the log of gdp per capita, and should be positive if economic development decreases the salience of the redistribution issue. Urbanization is the percent of the total population that lives in urban areas, and should have a negative coefficient if large concentrations of urban voters increases the salience of left-right redistribution among the poor. Excommunist is an indicator variable that takes the value 1 in ex-communist systems, and should have a negative coefficient if the poor are particularly sensitive to redistribution issues after the transition from communism to market economies. Polarization is the distance between the left-most right party and the right most left party on the economic dimension, using the Benoit and Laver data. The variable measures the emptiness of the center. As Polarization increases, voters on the left (right) must travel a greater ideological distance to support a party on the right (left). We should therefore expect this variable to have a negative coefficient. And Trade Openneness is the imports plus exports as a percentage of GDP. The coefficient should be negative if the uncertainties of globalization make poor workers seek insurance from the risk of open markets. Finally, we test the argument that poor voters will be most likely to vote for the economic right when the party system forces them to choose between economic redistribution and some other issue. We focus specifically on behaviors like abortion and homosexuality that have to do with non-economic individual behavior with negligible material externalities. The policy issues entail a trade-off between individual freedom of choice and state enforcement of specific conceptions of morality. No statist conservative party is an indicator variable that takes the value of 1 if a countries party system has no parties that are at once to the left on the economic dimension and to the right on Benoit and Laver s second dimension, Social. The social dimension ranges from 1 (party favors liberal policies on matters such as abortion, homosexuality and euthanasia ) to 20 (party opposes liberal policies on these issues). The 19

21 variable should have a positive coefficient because for poor voters to vote conservatively on issues of individual liberty, they must support parties that are also right-wing on redistribution. We also include control variables related to economic retrospective voting. If voters support the incumbent when the economy is performing relatively well, and reject them when the economy is performing relatively poorly, then voting polarization by income should be relatively low. Regardless of which party is in power, there should be support for the incumbent only if the economic performance so warrants. Voting support for the right should therefore be large when the incumbent government is on the right, and the economy is performing well. To test for economic voting, we include an indicator variable, Right incumbent, that takes the value of 1 if the left-right position of the government (weighted by the size of government parties) is to the right of Growth change is the change in GDP growth between the election year and the previous year, and Right incumbent*growth change is an interaction of these two variables. In the robustness checks, we also use two different measures of economic performance of the incumbent. Unemployment change is the change in unemployment between the election year and previous year, and Right incumbent*unemployment change is the interaction. Inflation change is the difference between the consumer price index inflation rate in the election year and its previous year, and Right incumbent*inflation change is the interaction. The coefficients on the first two interaction terms should be positive and the third should be negative if voters support right wing incumbents when the economy is doing well. Results We now turn to the second-level regressions. The dependent variable is the coefficient on Poor from the first-level regressions in each country ( b_poor ), which measures the difference in the probability that, all else equal, a poor person votes for a right wing party and a non-poor votes for a right wing party. Therefore, negative values of coefficients in the second-level regressions 20

22 imply more voting polarization by income, and positive values imply less voting polarization by income. Column 1 presents the kitchen sink model, where all of the independent variables discussed above are included. This model has a relatively poor overall fit (the adjusted R-squared is 0.58), and a number of coefficients are very imprecisely estimated. These include the controls for incumbency (using income growth as a measure of economic performance), the log of GDP, and the indicator for the existence of an extreme anti-immigration party, all of which have p- values greater than.75. Across a wide range of specifications which we do not report here, the results were consistently the same for these variables. They always have coefficients with very large standard errors that are never remotely close to conventional levels of statistical significance. The data therefore suggest that arguments about economic development, such as post-materialism or the decline in class voting, do not help us to understand the differences in support for right-wing parties across income groups. We also do not find evidence of economic voting by the poor in our analysis. And we find no evidence that the propensity of the poor to vote for right-wing parties on redistribution is affected by the presence of extreme antiimmigration right-wing parties. Column 2 reports the results when we include only those variables that are consistently good predictors of b_poor. Each of the coefficients on these variables has the expected sign and is precisely estimated in Columns 1 and 2. Polarization, for example, is negative and statistically significant, with t value respectively equal to 2 (in the more complete specification) and 4.2 in the parsimonious specification. The negative coefficient supports the argument that the cost of switching to an economically conservative party affects the trade-off faced by poor voters when for other reasons (e.g., to punish a bad incumbent, or because of a salient second dimension) they want to switch to a right party. When the cost is higher, cross-over is less likely to take place. This same claim is also supported by the negative coefficient on Excommunism. 21

23 Ethnic fractionalization has a positive coefficient, supporting the negative exposure argument, that poor voters will vote more like non-poor voters in countries with an ethnically diverse population. Religious fractionalization, by contrast, has a negative coefficient, implying that in religiously fractionalized countries, voting polarization by income is greatest when church power is lowest. Similarly. False witnessers, has a positive coefficient, suggesting that materially powerful churches affect the behavior of voters, making poor people more likely, all else equal, to support the same party that non-poor people support. Urbanization has a negative coefficient supporting the claim urbanization affects the values of voters, making the second (social) dimension coincide with the income redistribution dimension, hence reducing cross-over. Trade openness has a negative coefficient, supporting the conventional wisdom that the working class in more open economies demands the government to institute safety nets that shield it from the costs of price and demand oscillations in the international markets, making the poor are less likely to support the same party as the rich. Finally, the positive coefficient on No statist conservative party in the electoral arena supports the claim that cross-over to the economic right is more common when poor voters who have conservative preferences on a second individual rights dimension must vote for a party that is right-wing on redistribution in order to vote for their conservative values on individual liberties. Below, we interpret the substantive importance of these effects. Robustness. In order to check the robustness of the results reported in column 2 of Table 1, we estimate several different specifications of the model: we add further controls, we reestimate the parsimonious model with a different definition of right wing in the first level estimation and a different specification of the first level equation; and we re-estimate the more complete model using two alternative potential measures of incumbency and accountability factors. The results are in Table 2. 22

24 First consider some additional control variables. One could claim that our results regarding religious fractionalization and the network value of churches are mainly driven by the political activism of the Catholic Church in some European countries, hence we re-estimate the model including a control for the proportion of Catholics in a country. The coefficient (see column 1, Table 2) is not significant. One could also claim that the result regarding the higher polarization of the electorates in the former planned economies is due not to their past political economy and their path of increasing income inequality, but to the growing pains of younger democracies, hence we include a control for the age of the democratic regime, whose coefficient, as one can observe in Column 2, is not statistically significant. Finally, a history of state regulation of religion might lead to a less powerful role of the church and of religion in society, which should lead to greater voting polarization; the opposite claim that the political power of the church is better captured by its close relationship with the state apparatus rather than by its network value might also be made. Column 3 includes an indicator variable that takes the value 1 if there was an official state religion in The coefficient is not significant. Importantly, the inclusion of none of these variables affects the results for the other variables. The results described in the previous subsection are based on a particular specification of the first-stage model. The first level dependent variable takes the value of 1 if the party voted by the respondent scored above 10.5 on the economic dimension in the Benoit and Laver data. The first-level specification also includes only the Poor dummy variable to measure income, so the coefficient b_poor measures the propensity of a poor voter compared to a non-poor voter to vote for a right wing party. We re-therefore test the robustness of our results by estimating the model in Column 2 of Table 1 using different specifications of the first-stage regression. First, we adopt a stronger definition of right-wing: a party is considered right wing if it scores 12.5 or above on the economic redistribution dimension in the Benoit and Laver (2006) dataset. The coefficient b_poor still measures the propensity of the poor, versus the non-poor, to vote for a 23