The Politics of Inequality and Partisan Polarization in OECD Countries. Jonas Pontusson 1 and David Rueda 2

Transcription

1 The Politics of Inequality and Partisan Polarization in OECD Countries Jonas Pontusson 1 and David Rueda 2 Paper prepared for presentation at the annual meeting of the American Political Science Association. Washington DC, Sept. 1-4, Department of Politics, Princeton University, jpontuss@princeton.edu. 2 Department of Politics and International Relations and Merton College, University of Oxford, david.rueda@politics.ox.ac.uk.

2 2 Inequality has emerged as a major theme in the comparative studies of advanced capitalist political economies over the last decade. Much of the literature to date treats inequality as a dependent variable, addressing questions like the following: Why has wage inequality increased more in some countries than in others (Wallerstein 1999, Rueda and Pontusson 2000)? How do wage inequality trends and changes in the distribution of employment jointly affect the distribution of income among households (Kenworthy and Pontusson 2005)? And to what extent have welfare states compensated for rising market inequality (Bradley et al 2003, Kenworthy and Pontusson 2005)? In addressing such questions, political scientists typically invoke political institutions and government partisanship as explanatory variables. In this paper, we seek to break new ground by turning things around, i.e., by exploring the consequence of changes in income distribution for politics specifically, for party politics and electoral competition. To the extent that students of comparative political economy have broached the question of the political consequences of inequality, they have done so through the prism of median-voter theory. The standard prediction of median-voter theory is that rising inequality makes the median voter more prone to support redistributive policies and that rising inequality will therefore be associated with more redistribution (Romer 1975, Meltzer and Richard 1980). Several recent contributions (notably Moene and Wallerstein 2001, Iversen and Soskice 2001) have observed that the predicted association between inequality and redistribution does not hold crossnationally, at least not among OECD countries, and have offered alternative models of how inequality affects aggregate demand for social insurance and redistribution. The policy preferences of the median voter play an important role in what follows, but our analysis focuses on how inequality affects the preferences of core constituencies of parties of the Left and the Right and, as a result, the policy positions adopted by parties of the Left and the Right. In posing the question of the political consequences of inequality in this manner, we build on and extend the comparative political economy literature on partisan effects on social spending

3 3 and macro-economic policy (e.g., Hibbs 1987, Garrett 1998). Our approach also connects to McCarty, Poole and Rosenthal s (2004) analysis of the recent polarization of American politics. McCarty, Poole and Rosenthal document that partisanship in congressional roll-call voting declined in the 1950s, held steady through most of the 1960s and 1970s, and then increased sharply from the late 1970s onwards. This pattern parallels trends in income distribution in a very striking manner. Our goal is to determine whether a similar pattern of association between income inequality and partisan polarization holds for other OECD countries as well and to explore causal mechanisms that might account for such a pattern. Most of the comparative literature on partisan effects focuses on the question of what parties do when they are in government. Encompassing observations from twelve OECD countries over the period , our analysis instead measures partisanship by the platforms on which parties compete in elections--what they promise to do if they are elected. Using the Party Manifesto Dataset, we measure partisan polarization as the standard deviation of the positions on the Left-Right dimension of all parties that obtained parliamentary representation in a particular election. In calculating the standard deviation, we transform each party's score on the Left-Right dimension to account for its share of the vote in the election in question. Movements to the Left or the Right by large parties thus have a bigger impact on our measure of polarization than movements by small parties (or the emergence of new small parties on the far Right or the far Left). In other words, our measure of polarization is designed to capture polarization in the electorate as well as polarization among parties. We are interested not only in how inequality affects the spread of party positions on the Left-Right dimension, but also in how it affects the positions of parties of the Left, the Center and the Right. As measured by the standard deviation, political polarization might increase because parties of the Left and the Right move in opposite directions. However, we would also observe increased polarization if all parties move to the Right so long as the parties located on the Right move more sharply to the Right than parties located on the Left. To explore the effects of

4 4 inequality on party positions, we calculate (vote-weighted) mean scores on the Left-Right dimension for parties classified as Left, Center and Right by Swank in the codebook for his Comparative Parties Dataset. 3 In estimating the effects of inequality on party positions as well as party-system polarization, we control for a battery of other variables that might plausibly influence these outcomes (to be presented below). We present two sets of results, with and without fixed effects estimated with country dummies. While models without fixed effects address both variance across countries and variance over time, models with fixed effects speak primarily to variation over time. Put differently, our fixed-effects models hone in on the question of how parties (and voters) responded to changes in inequality over the period , bracketing the question of whether levels of inequality and partisanship were associated on a cross-national basis at the beginning of this period. To anticipate, our main findings can be summarized as follows. With and without fixed effects, more inequality is strongly associated with more partisan polarization. Without fixed effects, inequality is associated with Left parties being more leftist in their orientation and with Right parties--to a lesser extent also Center parties--being more conservative in their orientation. With fixed effects, inequality is still associated with Center-Right parties being more conservative, but we no longer observe an association between inequality and leftism on the Left. Our fixed-effects results actually indicate that Left parties have become more conservative as a result of rising inequality, but the Australian case turns out to be highly influential in this regard. Without Australia, there is no statistically significant association between inequality and the orientation of Left parties. In the course of the following discussion, we elaborate theoretical reasons why these results make sense. 3

5 5 1. Inequality and Partisan Competition As indicated above, the conceptual framework underpinning our inquiry is different from the median-voter framework. Median-voter models assume that parties are more or less entirely motivated by winning elections and have no enduring commitment to particular policies (or constituencies). In a two-party system, winning elections requires winning the support of the median voter. Thus parties should converge on the redistributive preferences of the median voter, in their election manifestos as well as their actual behavior in government. Although this logic may be less pronounced in multi-party systems, the median-voter framework posits that the effects of inequality on the policies advocated and pursued by parties operates though the policy preferences of the median voter. The alternative approach adopted here posits that parties of the Left and the Right have core constituencies to which they are historically and ideologically committed as well as organizationally tied. More or less formal union-party ties render Left parties particularly responsive to the interests of low-income earners. Right parties are particularly responsive to high-income earners by virtue of their traditional ties to business associations and professional organizations. As Powell (1982: 116) argues, the existence of a relationship between strong, continuing expectations about parties and the interests of social groups not only creates easily identifiable choices for citizens, it also makes it easier for parties to seek out their probable supporters and mobilize them at election time. But historical, ideological and organizational commitments are not enough: elections need to be won and they inevitably revolve around issues that give political meaning to partisan attachments and social divisions (Dalton 2002: 195).

6 6 There every reason to suppose that the income of the median Left-party voter is lower than the mean income while income of the median Right-party voter is higher than the mean income in everyone of the countries included in our analysis. Rising inequality means that the distance between the income of the median voter for both parties and the mean income of all voters increases. A partisan version of the standard median-voter argument (Romer 1975, Meltzer and Richard 1980) readily suggests itself: as inequality rises, the median Left-party voter should want more redistribution and the median Right-party voter should want less redistribution than what they previously wanted. By this logic, we expect inequality to be associated with partisan polarization. We do not mean to suggest that parties are oblivious to the preferences of the median voter in the electorate as a whole. Following Strom (1990) and Garrett (1998), among others, we assume that parties are motivated by winning elections and, at the same time, by serving the interests of their core constituencies. In government, parties can be expected to pursue partisan distributive objectives so long as these policies do not threaten their prospects of re-election. In a similar vein, Aldrich (1983, 1995) argues convincingly that parties need party activists and that the parties median voter may be equally or more influential than the median voter in the electorate as whole. Again, it is reasonable to suppose that rising inequality affects party strategies through its influence on the preferences of the median voter of each party. The distribution of income is surely not the only thing that matters for the policy preferences of core constituencies, let alone the actual policy positions adopted by parties. There are a number of compelling reasons to believe that economic and political conditions since 1980 have constrained the ability of governments to pursue redistributive policies. To begin with, one might well argue that redistributive welfare states have grown to their limits. Especially in the context of slow economic growth, tax fatigue among middle-income voters appears to have become an increasingly potent constraint on further expansion of redistributive spending. Secondly, pressures associated with globalization capital mobility in particular might be said

7 7 to represent constraints on redistributive policy. The extent to which globalization in fact constrains redistributive policy has been a topic of considerable debate among students of comparative political economy, but there can be little doubt that politicians of the Left and the Right alike perceive globalization as a constraint. Thirdly, it seems most plausible to argue that declining union membership and mobilizational capacity has resulted in lower rates of electoral turnout among low-income workers in many OECD countries. Holding the distribution of income constant, declining voter turnout at the lower end of the income distribution implies that the income of the median voter improves relative to the mean income, and hence the median voter becomes less supportive of redistribution. Like a number of other analysts before us, we also recognize that there are both centripetal and centrifugal forces at work in electoral competition and that the relative magnitude of these centripetal and centrifugal forces varies across types of electoral and party systems (see, for example, Cox 1990 and Grofman, 1993). The number of parties in a system shapes the incentives for either centripetal patterns of competition (where parties have strong incentives to move to the center of the ideological spectrum) or centrifugal ones (where parties have more flexibility in their choice of strategies). In estimating the effects of inequality on party positions on the Left-Right dimensions, our empirical models control for the effective number of parties. We also control for several political-economic variables that might plausibly constrain the ability of parties to respond to changes in the redistributive preferences of their core constituencies; in particular, variables that might constrain the ability of Left parties to respond to greater demand for redistribution among unskilled workers. In addition, we include Kim and Fording s estimates of the position of the median voter in our models of partisan polarization as well as our models of the (vote-weighted) mean positions of party families. Using the Party Manifesto Dataset, Kim and Fording compute the position of the median voter on the Left-Right dimension from the distribution of the

8 8 electorate among parties with distinct positions on this dimension (see Kim and Fording 2002, 2003 for details). Their measure ranges from 0 (Left) to 100 (Right). Averaging across the twelve countries included in our analysis, Figure 1 traces the evolution of the position of median voter on the Left-Right dimension, as measured by Kim and Fording, over the last three decades of the twentieth century. After an initial dip in the early 1970s, when the median voter became more left-leaning, we observe a steady increase in the conservatism of the median voter. The median voter goes from a position close to 40 in the ideological spectrum in the early 1980s to one close to 60 in the late 1990s. Considering that income inequality increased in virtually all OECD countries in the 1980s and 1990s (see Rueda and Pontusson 2000, Kenworthy and Pontusson 2005), this evidence would seem to fly in the face of the standard Meltzer-Richard argument. 4 Again, however, our analysis is not directly concerned with the effects of inequality on the preferences (or ideological position) of the median voter. 5 [Figure 1] The evidence presented in Figure 1 suggests that Left parties and Right parties faced very different situations in the 1980s and 1990s. Our argument about the interests of core constituencies holds that rising inequality put pressure on Left parties to adopt a more redistributive orientation while it put pressure on Right parties to adopt a less redistributive orientation. At the same time, the preferences of the median voter became distinctly less redistributive. For Right parties, the need to maintain support among core constituencies and the need to appeal to the median voter pointed in the same (rightward) direction. By contrast, Left parties have found themselves in a more contradictory or ambiguous situation as the preferences 4 Further documentation of rising income inequality will be included in the next iteration of this paper. 5 It should be noted that our preliminary analyses, including a number of control variables, do not yield any statistically significant association between wage inequality and the position of the median voter (results available upon request). The bivariate correlation between wage inequality and the Left-Right placement of the median voter in our data is only.31.

9 9 of their core constituencies and the preferences of the median voter have appear to have moved apart. 2. Measuring party positions and partisan polarization Our measures of the ideological positions of parties and the degree of partisan polarization in any given country-year are based on the Left-Right index developed by the Comparative Manifestos Project. The CMP codes the policy emphases of election programs under a large number of separate categories. Fifty-six of these categories are summarized in a Left-Right index, ranging from the extreme Left, -100, to the extreme Right, As documented by Gabel and Huber (2000), the index values generated by this procedure correlate reasonably well with various party classification schemes based on expert surveys (see also McDonald and Kim n.d.). Moreover, several of studies (e.g., Powell 2000) have shown both that the Left-Right dimension is a good summary of what parties stand for in elections and that it is a meaningful factor for voters and there is also some evidence that there is a correlation between Left-Right positions as measured by CMP and various policy outcomes (e.g., Budge and Hofferbert 1990). 6 Using the Left-Right dimension of the CMP as the basis for testing the arguments set out in the previous section obvious assumes that preferences over redistribution constitute a core component of the Left-Right dimension. This strikes us as a rather reasonable assumption. It should be noted there are not a lot categories in the Manifesto Dataset that deal specifically with redistribution. The only clear candidate would be the variable that measures a welfare state dimension. Measuring the emphasis that the welfare state receives in party manifestos is, in our 6 For more information about the CMP data, see Budge et al 1987 and Laver and Budge 1992.

10 10 view, too limited and potentially misleading. As the existing welfare-state literature emphasizes, the advocacy of the public provision of social welfare may be motivated by a number of considerations other than redistribution. In addition, we are not aware of any studies that explore how the welfare-state dimension correlates with expert judgements or its relevance for policy outcomes. For all these reasons, we prefer to use the Left-Right dimension as our proxy for preferences with respect to redistribution. Relative to party placements based on expert judgments, the big advantage of manifestobased party placements is that they change over time, which is obviously essential our present purposes. At the same time, it must be noted that the measures of party positions that we obtain from the Manifesto dataset exhibit a great deal of election-to-election volatility. In part, this volatility is probably a reflection of measurement error, but we believe that it also reflects what tactical maneuvering by parties and, specifically, signaling behavior. Clearly, voters do evaluate parties strictly based on their platform (pledges) in the current election. Most voters know something (perhaps a great deal) about what parties have said and done in the past. As a result, parties are likely to rhetorically exaggerate changes in their policy commitments, to make sure that voters recognize such changes. The true position of parties, therefore, is one that requires us to discount some of the volatility that we observe in the data. Following the lead of MacDonald and Mendes (n.d.), we do this by transforming the ideology score to be the average of the present and the two previous elections. The resulting measure is highly correlated with the original (i.e., we still capture the essence of the original measure), but it is much smoother and less volatile on an election-to-election basis. As we mentioned at the outset, we measure political polarization as the standard deviation of the Left-Right dimension of all parties that obtain parliamentary representation in a particular election, taking into the share of the vote received by each party. In measuring polarization in this fashion, we conform to the dominant approach in the party-systems literature

11 11 (Taylor and Herman 1971, Sigelman and Yough 1978, Hazan 1995, Knutsen 1998). 7 To calculate polarization, we first take each party's score on the Left-Right dimension and multiply it by its share of the vote (measured as a fraction) in the election in question. This new variable ranges from -100 (an extreme Left party that obtains 100 percent of the vote) to +100 (an extreme right party that obtains 100 percent of the vote). One way to think about the transformation of the ideology measure is that the positions in the Left or the Right of large parties have a bigger impact on our measure of polarization than those of small parties. Not giving the same importance to small and large parties seems reasonable. Another way to think it is that our measure of polarization is designed to capture polarization in the electorate, as well as polarization among parties. To do so, we need to know how widely-held the different ideological positions are in a particular election. The second step in our measure of polarization is to calculate the standard deviation of the new ideology variable considering every party that obtains parliamentary representation. 8 The standard deviation is a measure of the degree of dispersion around the mean and therefore a natural candidate for assessing polarization. Because the standard deviation is the root mean square deviation from the average, it is an intuitive measure of polarization. In a simple sense, it is the average dispersion around the mean measured in the same units as the initial variable. If our polarization measure is 10, for example, that will mean that the average dispersion around the mean in a particular country and a particular election is 10 units in the transformed Left-Right continuum. In addition to being interested in the determinants of polarization as an aggregate measure, we want to explore the specific effects of inequality on the positions of parties of the Left, the Center and the Right. Again, polarization as measured by the (vote-weighted) standard 7 Note that the aforementioned authors measure polarization by the variance of vote-party positions rather than the standard deviation. There is no reason to believe that this specification issue has any substantive implications. 8 In some countries, parties that get an insignificant portion of the popular vote will not be included in the manifesto dataset. See the Comparative Manifesto documentation for details.

12 12 deviation might be the result of several different party movement scenarios. We are particularly interested in exploring the proposition that Left parties have been more constrained than Right parties in responding to changes in the redistributive preferences of their core constituencies. To assess the validity of this proposition, we calculate the mean scores on the Left-Right dimension for parties classified as Left, Center" and Right by Swank, with individual party scores again being transformed to take into account the percentage of the votes received by each party. 9 It is important to point out that for all three of these measures, higher values mean more conservative positions while lower values mean more liberal ones. Figure 2 shows the yearly averages for the polarization data as well as for the ideology positions of Left, Center and Right parties since The solid blue line represents polarization. Polarization is measured as standard deviation, as explained above, as it is therefore the average dispersion around the mean for each year. As mentioned before, this variable is measured in units of transformed ideology and it is always positive. The dashed red line is the position of the Left. This is the ideological mean (transformed to account for the percentage of vote) for all parties in the Left each year. The dashed green line is the position of the Right and the dashed yellow line is the position of the Center. As a reminder: the Left, Center and Right variables are ideological positions going from -100 (extreme Left with 100 percent of the vote) to +100 (extreme Right with 100 percent of the vote). [Figure 2] Figure 2 makes clear that there has been no OECD-wide secular increase in polarization. The polarization variable exhibits a high degree of variation through time but no secular trends. As for the mean positions of families of parties, Figure 2 indicates that there has been a general movement to the right by both Center and Left parties. We must keep in mind, however, that these are averages for all OECD countries. Country-specific graphs (not presented but available 9 See Appendix 2 for the classification of the parties used here. Note that this classification is not exhaustive: for some instances, our measure of polarization includes small parties that are not included in our measures of the positions of party families (Left, Center and Right).

13 13 from the authors) show that there is a wide degree of variation across countries with respect to party movements on the Left-Right continuum as well as partisan polarization. The more systematic analysis to be developed in the following pages will allow us to ascertain whether these developments are related to changes in the levels of inequality. 3. Measuring inequality There are two possible sources of comparative measures of income inequality in OECD countries: the Luxemburg Income Study, which pertains to the earnings/income of households, and the OECD dataset on relative wages among full-time employees. Arguably, the distribution of income among households is more relevant to the political preferences of voters than the distribution of wages among individuals, but the LIS data on income inequality are far more fragmentary than the OECD data on wage inequality. For disposable income inequality, LIS yields a total of about 70 country-year observations that we might use and for market income inequality (before taxes and transfers), this figures drops below 60. As a first cut, the following analysis relies on the OECD wage dataset on relative wages. Specifically, our measure of inequality is the wage ratio, i.e., the ratio of earning at the 90th percentile to earnings at the 10th percentile of the wage distribution. A measure of the distance between two points, the ratio certainly does not tell us everything that we would want to know about the overall shape of the distribution, but it is a commonly used measure and easy to interpret. Higher values signify greater levels of inequality. The reader should keep in mind that our inequality measure ignores important sources of income, such as self-employment, income from capital, and government transfers. It also ignores the distributive effects of taxation and income pooling within households. Moreover, the OECD dataset on which we rely is restricted to full-time employees (except in the case of Austria).

14 14 What follows, then, must not be confused with an analysis of the overall distribution of income in OECD countries. This said, income from employment accounts for the lion s share of income in all OECD countries and the distribution of income from employment, as measured by ratios, correlates quite closely with broader measures of income distribution on a cross-national basis (see OECD 1995, Gottschalk and Smeeding 1997, Kenworthy and Pontusson 2005). Where wage inequality data were missing in the middle of a time series, we engaged in linear interpolation. With 9 interpolated data points, this procedure yielded a total of 179 observations covering 12 countries over the period 1970 and (The countries included in the sample are: Australia, Denmark, Finland, France, Germany, Italy, Japan, the Netherlands, Sweden, Switzerland, the UK and the US). It should be noted that the country panels are quite unbalanced: the shortest time series is 3 years while the longest is 26 years. The average for all countries is 15 years. To clarify, country-years constitute our units of observation. For each country-year, our manifesto-based measures of party positions and partisan polarization refer to the first election after the year that we have wage inequality for. Thus wage inequality in non-election years is matched with partisanship in the subsequent election (and, again, partisanship scores are an average for that and the preceding two elections). 4. Control variables In exploring the effects of inequality on the ideological positions and electoral strategies of political parties, we want to control for the effects a number of other variables that might be expected to affect what parties promise in their election platforms. Let us briefly introduce these control variables, which are drawn from the existing literatures on party systems and comparative

15 15 political economy. As indicated below, we do not have strong theoretical expectations about the effects of these variables. 10 Effective number of parties. Among others, Cox (1990) and Grofman (1993) argue that the number of parties in a system shapes the incentives for either centripetal or centrifugal patterns of competition. To explore such effects, we use the measure developed by Laakso and Taagapera (1979) and coded by Armingeon et al (2004). GDP growth. Economic stagnation may be a source of partisan polarization. In response to slow growth, Center and Right parties are likely to pursue more Right-leaning policies. The effects of slow growth on the partisan strategies of Left parties are more ambiguous. International and Financial Openness. There are two contradictory accounts of the effects of internationalization on partisan politics. Generally, as argued by Wood (1994), some dimensions of international openness are believed to promote inequality in industrialized democracies. The connection with partisan politics, however, is more contentious. There is first a large literature suggesting that growing levels of international openness, integration and interdependence result in a blurring of partisan differences caused by the inability of social democratic parties to produce policies that do not conform to market forces (see, for example, Iversen 1996 and Scharpf 1991). Then there are some authors who argue either that international forces do not affect some partisan differences (like Boix 1998 and Garrett and Lange 1991) or that they actually have strengthened the influence of partisanship on policies and economic outcomes (Garrett 1998). International openness is measured both as imports and exports in percent of GDP and as the sum of several indexes for financial restrictions (for details, see Armingeon, Beyeler and Menegale 2004). 10 Appendix 1 presents summary statistics for all the variables used in our analysis.

16 16 Union density. The power and strategies of unions are most influential in the ability of Left, Center and Right parties to move to the Right or to the Left. The union density measure used in this paper represents employed union members as a percentage of employed labor force. The figures were taken from Ebbinghaus and Visser (2000) except for Australia, Japan, the UK and the US (pre-1990 figures were taken from Visser 1996 and post-1990 figures provided by Bernhard Ebbinghaus). Social Security Transfers. As we mentioned above, it is possible that the size of the welfare state (maybe having reached its limits) influences the ideological position of parties. Can Left parties move to the Left when the welfare state has reached full maturity? To explore these issues, we introduce a measure of social security transfers as a percentage of GDP into our analysis. This variables consists of benefits for sickness, old-age, family allowances, etc., social assistance grants and welfare benefits paid by general government. (For details, see Armingeon, Beyeler and Menegale 2004). Median voter. As noted above, we are primarily interested in exploring how parties react to the demands of their core constituencies, but in so doing we wish to control for the position of the median voter. The measure in our analysis assesses the position of the median voter looking at the distribution of the electorate in regards to the Left-Right position of the parties in each election (for details, see Kim and Fording 2002 and 2003). Using party scores from the Manifesto Dataset, our measure of the median voter s position on the Left-Right dimension is computed as the median in grouped data determined from the vote shares of the parties contending an election, with higher values representing more conservative positions.

17 17 5. Statistical models For the empirical analysis, the model specification is as follows: Y it = β 0 +β 1 X 1it + +β n X nit +ε it where β 0 represents a general intercept, X 1 to X n are the explanatory variables, β 1 to β n are the slopes of the explanatory variables, and ε it denotes the errors. We performed a modified Wald test for panel-specific heteroscedasticity. This test revealed a significant amount of heteroscedasticity at the level of the countries. We therefore present all our results with panel-corrected standard errors (PCSEs). Beck and Katz (1995) show that, in the absence of autocorrelation, PCSEs are consistent when there is panel-specific heteroscedasticity. 11 We present two sets of results. First, we report results without fixed effects. We do this because we are interested in assessing the relationship between inequality and polarization when we compare countries as well as years. In the analysis without fixed effects, in other words, we look at all the variance within the polarization variable (whether it is the result of comparing observations across countries or through time). We call this set of results our aggregate results. Second, we report results with country fixed effects. In this set of models, which include a constant and dummy variables for all countries but one (Australia serving as the reference country), we only look at the variance within countries. These results reflect the average withincountry relationship between inequality and polarization, but do not take into account of differences across countries.

18 18 6. Results I: Determinants of polarization Let us first look at the relationship between inequality and polarization. Table 1 presents our aggregate results. This is the analysis without country fixed effects and therefore the one that looks at all variance, across countries as well as over time. [Table 1] Table 1 makes clear that wage inequality is a highly significant determinant of polarization (significant, in fact, at better than the 99% level of confidence). A one-unit increase in the ratio is associated with an increase in polarization that equals more than 6 ideology units. As shown in Appendix 1, the mean for all countries for wage inequality is a ratio equal to 2.99, with a standard deviation of According to our results, an increase in inequality equal to one standard deviation (for example, from 2.99 to 3.61) is associated with an increase in polarization equal to 3.94 ideology points. This increase in polarization is larger than one standard deviation in polarization for all countries (see Appendix 1). The estimated effect of inequality on polarization would appear to be very significant from a substantive as well as a statistical point of view. Table 1 also reports the estimated effects of the control variables included in our model. Only two variables (GDP growth and Social Security Transfers) are not significant determinants of polarization in the OECD. According to these results, neither sluggish growth nor the limitations of the welfare state affect the ideological differences among parties. The number of effective parties is negatively associated with polarization. The coefficient is statistically significant but the substantive effects do not seem to be large. Each additional party is associated with a decrease of less than one ideology unit of polarization. Both measures of internationalization included in the analysis, on the other hand, are positively associated with polarization. The more international and open an economy, our results suggest, the more 11 The results of all tests are available from the authors.

19 19 polarized the party system. Increasing union density is also associated with more polarization. We would expect this variable to mainly work by pushing Left parties to the left and will explore this issue below. The position of the median voter is the last significant variable. Our results suggest that more conservative median voters are associated with more polarized party systems. Whether this variable s affects work mainly by pushing all parties to the Right (but to different degrees) is a question that we will also explore in the next section. To reiterate, the results presented in Table 1 speak to the determinants of cross-national and over-time variance in partisan polarization taken together. By estimating a fixed-effects model with country dummies, we focus on the consequences of changes in wage inequality for over-time variation in political polarization within countries. Table 2 presents these results (country-dummy estimates available upon request). [Table 2] As with Table 1, the most important thing to report about Table 2 is that wage inequality is a very significant determinant of polarization. Even when we eliminate cross-country variation from our estimation, higher levels of inequality are associated with over-time increases in polarization. This relationship is again significant at better than the 99% confidence level. This result is all the more remarkable since all other explanatory variables in the model are insignificant at better than the 95% confidence level. It is certainly the case that the substantive effect of inequality when looking only at variation over time is much smaller than we also looked at cross-country variation. Still, a one unit increase in the ratio is associated with an increase in polarization that equals more than 3 ideology units. With fixed effects, none of the other variables included in our model clear the 95% threshold of statistical significance. The only other variable that might be said to affect polarization through time is trade openness (the sums of exports and imports in percent of GDP). When looking at change over time and within countries, increasing internationalization appears to

20 20 be associated with greater polarization, but this association is only statistically significant at the 90% level of confidence. 7. Results II: Determinants of party positions As noted above, increased polarization might be a result of several different combinations of movements by Left, Center, and Rights parties in the Left-Right dimension. In this section, we analyze the determinants of party positions on the Left-Right dimensions. In Table 3, we present the results for the aggregate analysis (i.e., the analysis that considers both cross-country and overtime variation). There are three different models in Table 4: one for the average position of the Left parties, one for the average position of Right parties, and the third one the average position of Center parties. Again, our mean scores for the Left, Right and Center are vote-weighted, taking into account for the percentage of votes received by the different parties making up each bloc of parties, and higher scores always signify more Right-leaning policy positions. The total number of observations varies across the three models because not all countries have parties that we classify as Left, Center and Right at all times. Most notably, we follow the convention of classifying the American Democratic Party as a centrist rather than leftist party: consequently, no US observations were used to estimate the first of the models reported in Table 3. [Table 3] As hypothesized, the results in Table 3 indicate that higher levels of inequality are associated with more leftist Left parties and more rightist Right parties. The relationship between inequality and the positions of Left and Center parties is statistically significant at better than the 99% confidence level. When looking at cross-country and over-time variation, a one unit increase in the ratio is associated with a 5.6 unit movement to the Left by Left parties and with a 5.9 unit movement to the Right by Right parties. An additional interesting finding is that

21 21 Center parties in the OECD have responded to increasing inequality by moving to the Left. In substantive terms, however, the movement to the Left of Center parties is not as significant as that of Left parties--or as the rightward movement of Right parties. Several of the control variables display significant effects on the strategies of Left, Right and Center parties. In addition to wage inequality, the only variable in Table 3 that is a statistically significant determinant of the strategies of all three groups of parties is the effective number of parties. This variable has exactly the opposite effect of inequality. An additional party is associated with more conservative Left and Center parties but with less conservative Right parties. Economic growth only affects one group of parties in a significant way: it makes Left parties more conservative when growth is high or, conversely, more leftist when growth is low. Higher levels of internationalization, when measured as imports plus exports, move Left parties to the Left and Right parties to the Right but they have no effect on Center parties. When internationalization is measured as openness to capital mobility, however, it only affects Right parties (by making them more conservative). As one would expect, high union density is associated with more leftist Left parties but, surprisingly, it is also significantly associated with more conservative Right parties. It is also noteworthy that social security transfers are only significant as a determinant of the position of Right parties. More generous welfare states are associated with more conservative Right parties. There are no surprises as far as the effects of the position of the median voter are concerned: more conservative median voters invariable push all parties to the Right. This effect is significant at better than the 95% confidence level in the case of Left and Right parties, but only at the 90% level in the case of Center parties. Table 4 present the results of models identical to those in Table 4 except that the incorporate fixed-effects through country dummies. One striking finding immediately jumps out: when we restrict our analysis to over-time variation in party positions, increasing inequality is associated with more conservative positions adopted by all parties, i.e., parties of the Left (and

22 22 the Center) as well as parties of the Right. We would not expect Left parties to move to the right when facing increasing levels of wage inequality, but our results seem to suggest exactly this. When focusing on over-time variation within the countries in the sample, increases in wage inequality seem to make Left parties less leftist. This conservatizing effect is also present in the models with Right and Center positions as the dependent variables. It is important to point out that the results in Table 4 are entirely with those in Table 2 regarding partisan polarization. In Table 2, we found that wage inequality was associated with increasing polarization when looking at variation over time. The results in Table 4 suggest that this increase in polarization is due to the fact that all parties move to the right but Left parties do it at half the rate than Center and Right parties. [Table 4] The intriguing nature of the relationship between inequality and the ideology positions of Left parties warrants more attention. As noted at the outset, many reasons have been cited to suggest that the room for redistributive politics has diminished since the 1970s. Among other things, the options available to Left parties appear to have been constrained by lower economic growth, skills-biased technological change, production changes, the emergence of post-fordism, increasing internationalization, and competition from industrializing countries. 12 We have tried to control for these constraints. One way to interpret the association between inequality and more conservative Left party positions would be that our control variables do not fully capture the extent to which the political climate has become less favorable to redistribution. In a related vein, inequality may be associated with income-differentiate changes in voter turnout that affect the electoral incentives facing Left parties. Trying to capture this potential effect of inequality thus becomes an important task for us. It should also be noted that the estimate of the effects of inequality on the ideological position of Left parties reported in Table 4 is dominated by what we could call a country outlier:

23 23 Australia. If we repeat the analysis and eliminate a country at a time from our estimation the relationship between inequality and Left party positions disappears as soon as we exclude Australia. No other country is as influential on the estimates. The dramatic rightward turn of the Australian Labour Party appears to overwhelm the results in Table 4. In terms of the other explanatory variables, only one additional variable is significant as a determinant of the positions of the three parties: social security transfers. Our results suggest that when looking at variation through time increasing levels of social security transfers are significantly associated with more conservative position, no matter whether we look at Left, Right or Center parties. The position of the median voter is almost as significant. More conservative median voters move parties to the Right, but this association is only significant at better than the 95% level of confidence regarding the Left and the Right. For the Center, the statistical significance only reaches the 90% level. Internationalization, measured as openness, makes Left and Right parties more conservative but internationalization, measured as imports plus exports, only makes Right parties more conservative. Lastly, an increasing number of parties is associated with more conservative Center parties while higher union density promotes more leftist Left parties. 8. Conclusions In sum, all of our results are consistent with the proposition that inequality is an important source more partisan polarization. Taking cross-sectional as well as time-series variance into account, inequality is associated with Left parties being more leftist in their orientation and with Right parties--to a lesser extent also Center parties--being more conservative in their orientation. With fixed effects, however, inequality is no longer associated with more 12 We could include the increasing importance of insider-outsider differences in this list. See Rueda (2001 and 2005).

24 24 leftist parties: if rising inequality over the period had any consistent effect on the orientation of Left parties, that effect was a conservatizing one. Again, our findings underscore the constraints on the ability (or perhaps inclination) of Left parties to respond to increased demand for redistribution among workers at the lower end of the income distribution. With inequality rising while the center of political gravity has shifted to the Right, Left parties have been deeply cross-pressured. For Right parties, by contrast, the rightward movement of the median voter and the pressures of globalization have essentially reinforced the policy implications of inequality-induced changes in the redistributive preferences of core constituencies.

25 25 Table 1 The Relationship Between Wage Inequality (90-10 Ratio) and Polarization: Aggregate Results VARIABLE COEFFICIENTS (STANDARD ERRORS) P-VALUES Constant (2.609) Wage Inequality (90-10 Ratio) (.494) Effective Number of Parties (.118) GDP Growth (.075).301 Internationalization (Imports + Exports).057 (.011) Internationalization (Openness).258 (.099).009 Union Density.096 (.010) Social Security Transfers (.037).275 Position of the Median Voter.053 (.016).001 N 179 R 2.72 Notes: Results from OLS regression with panel-corrected standard errors. P-values are two-sided.

26 26 Table 2 The Relationship Between Wage Inequality (90-10 Ratio) and Polarization: Country Fixed-Effects Results VARIABLE COEFFICIENTS (STANDARD ERRORS) P-VALUES Constant (3.892).206 Wage Inequality (90-10 Ratio) (1.061).005 Effective Number of Parties.043 (.192).821 GDP Growth.032 (.068).638 Internationalization (Imports + Exports).043 (.024).076 Internationalization (Openness).128 (.126).309 Union Density.0001 (.035).996 Social Security Transfers.092 (.067).171 Position of the Median Voter (.014).136 N 179 R 2.84 Notes: Results from OLS regression with panel-corrected standard errors. P-values are two-sided.

27 27 Table 3 The Relationship Between Wage Inequality (90-10 Ratio) and Party Positions: Aggregate Results VARIABLE LEFT RIGHT CENTER Constant (3.557).019 Wage Inequality (90-10 Ratio) (.703) Effective Number of Parties.698 (.143) GDP Growth.247 (.107) (3.381) (.648) (.222).084 (.101) (3.552) (.807) (.137).049 (.102).626 Internationalization (Imports + Exports) (.013).051 (.015) (.009).258 Internationalization (Openness).219 (.161).174 Union Density (.013) Social Security Transfers (.046).950 Position of the Median Voter.055 (.024).020 N 158 R (.111).162 (.016).195 (.068) (.019) (.125) (.011) (.055) (.018) Notes: Results from OLS regression with panel-corrected standard errors. P-values are two-sided.

28 28 Table 4 The Relationship Between Wage Inequality (90-10 Ratio) and Party Positions: Country Fixed-Effects Results VARIABLE LEFT RIGHT CENTER Constant (5.468) Wage Inequality (90-10 Ratio) (1.401) Effective Number of Parties.058 (.298).846 GDP Growth.138 (.084) (4.491) (1.039) (.245) (.071) (5.303) (1.446).507 (.160) (.063).140 Internationalization (Imports + Exports).036 (.034) (.030).055 (.021).010 Internationalization (Openness).312 (.152).040 Union Density (.036) Social Security Transfers.333 (.090) Position of the Median Voter.042 (.019).024 N 158 R (.125) (.039) (.076).060 (.019) (.086) (.043) (.057) (.013) Notes: Results from OLS regression with panel-corrected standard errors. P-values are two-sided.