Brahmin Left Versus Merchant Right: Changing Political Cleavages in 21 Western Democracies, 1948-2020

This article sheds new light on the long-run evolution of political cleavages in 21 Western democracies. We exploit a new database on the socioeconomic determinants of the vote, covering more than 300 elections held between 1948 and 2020. In the 1950s and 1960s, the vote for social democratic, socialist, and afﬁliated parties was associated with lower-educated and low-income voters. It has gradually become associated with higher-educated voters, giving rise in the 2010s to a disconnection between the effects of income and education on the vote: higher-educated voters now vote for the “left,” while high-income voters continue to vote for the “right.” This transition has been accelerated by the rise of green and anti-immigration movements, whose distinctive feature is to concentrate the votes ∗ We are grateful to Luis Bauluz, Carmen Durrer de la Sota, Fabian Kosse, Marc Morgan, and Alice Sodano for their help in building the data set exploited in this article. We thank the editors, Nathan Nunn, Lawrence Katz, and An-drei Shleifer; six anonymous referees; Thomas Blanchet, Lucas Chancel, Ignacio Flores, Javier Padilla, Tom Raster, and Till Weber; and seminar and conference participants from the Paris School of Economics Applied Economics Lunch Sem- inar and the CUNY Graduate Center Comparative Politics Workshop for help-ful comments. We acknowledge ﬁnancial support from the Agence Nationale de la Recherche under the framework of the Investissements d’avenir programme reference ANR-17-EURE-001. of the higher-educated and lower-educated electorates. Combining our database with historical data on political parties’ programs, we provide evidence that the reversal of the education cleavage is strongly linked to the emergence of a new “sociocultural” axis of political conﬂict. JEL Codes: D72, P16, P51.


I. INTRODUCTION
Western democracies have undergone major transformations in recent years, embodied by political fragmentation, the increasing salience of environmental issues, and the growing success of antiestablishment authoritarian movements (Trump, Brexit, Le Pen, etc.). Yet, much remains to be understood about the nature and origins of these political upheavals. On what dimensions of political conflict (education, income, age, etc.) have such transformations aligned? Is the rise of "populism" the outcome of recent trends (such as the 2007-2008 financial crisis, immigration waves, or globalization), or can we trace it back to longer-run structural changes? Beyond country-specific factors, can we find evolutions that are common to all Western democracies?
This article attempts to make progress in answering these questions by exploiting a new data set on the long-run evolution of electoral behaviors in 21 democracies. Drawing on nearly all electoral surveys ever conducted in these countries since the end of World War II, we assemble microdata on the individual determinants of the vote for over 300 elections held between 1948 and 2020. Together, these surveys provide unique insights into the evolution of voting preferences in Western democracies. The contribution of this article is to establish a new set of stylized facts on these preferences and explore some mechanisms underlying their transformation in the past decades. 1 Comparing the evolution of electoral cleavages requires grouping political parties in such a way that the coalitions considered are as comparable across countries and over time as possible. To do so, we start by making a distinction between two large groups of parties: social democratic, socialist, communist, and green parties ("left-wing" or "social democratic and affiliated" parties) on one side, and conservative, Christian democratic, and anti-immigration parties ("right-wing" or "conservative and affiliated" parties) on the other side. 2 The most relevant result that emerges from our analysis is the existence of a gradual process of disconnection between the effects of income and education on the vote. In the 1950s-1960s, the vote for social democratic and affiliated parties was "class-based," in the sense that it was strongly associated with the lower-income and lower-educated electorate. It has gradually become associated with higher-educated voters, giving rise in the 2010s to a divergence between the influences of income (economic capital) and education (human capital): high-income voters continue to vote for the right, while high-education voters have shifted to supporting the left. This separation between a "Merchant right" and a "Brahmin left" is visible in nearly all Western democracies, despite their major political, historical, and institutional differences. 3 We also find that the rise of green and anti-immigration parties since the 1980s-1990s has accelerated this transition-although it can only explain about 15% of the overall shift observed-as education, not income, most clearly distinguishes support for these two families of parties today.
As a result, many Western democracies now appear to have shifted from "class-based" to "multidimensional" or "multiconflictual" party systems, in which income and education differentially structure support for competing political movements. One might call these systems "multi-elite" party systems, in which governing coalitions alternating in power tend to reflect the views and interests of a different kind of elite (intellectual versus economic), 2. We include parties commonly classified as liberal or social-liberal in this latter group, such as the Liberal Democrats in Britain and the Free Democratic Party in Germany. In Section II.B, we perform several robustness checks to ensure that our classification is consistent in terms of parties' programmatic supply and voters' own perceptions of the political space.
3. In India's traditional caste system, upper castes were divided into Brahmins (priests, intellectuals) and Kshatryas/Vaishyas (warriors, merchants, tradesmen), a division that modern political conflicts in Western democracies seem to follow to some extent. assuming that elites have a greater influence on political programs and policies than the rest of the electorate. 4 To shed light on the factors underlying the divergence of the effects of income and education on the vote, we match our data set with the Comparative Manifesto Project database, the most comprehensive available data source on the evolution of political parties' programs since the end of World War II. Drawing on two indicators of party ideology from the political science literature (Bakker and Hobolt 2013), corresponding to parties' relative positions on an "economic-distributive" axis and a "sociocultural" axis, we provide evidence that the separation between these two dimensions of political conflict and the divergence of income and education are tightly related phenomena. Specifically, we document that the correlation between parties' income gradient and their position on the economic-distributive dimension has remained very stable since the 1960s: parties emphasizing "profree-market" issues receive disproportionately more votes from high-income voters today, just as they used to 60 years ago. Meanwhile, the correlation between the education gradient and parties' positions on the sociocultural axis has dramatically increased over time, from 0 in the 1960s to nearly 0.5 in the 2010s.
In other words, parties promoting "progressive" policies (green and traditional left-wing parties) have seen their electorate become increasingly restricted to higher-educated voters, while parties upholding more "conservative" views on sociocultural issues (anti-immigration and traditional right-wing parties) have concentrated a growing share of the lower-educated electorate. We also find a strong and growing cross-country association between ideological polarization on sociocultural issues and the reversal of the education cleavage. In particular, the two countries in our data set where this reversal has not yet occurred, Portugal and Ireland, are also where partisan divisions over these issues remain the weakest today. Taken together, these results suggest that changes in political supply, in particular the increasing emphasis on sociocultural factors among old and new parties, appear to be an important factor behind the progressive disconnection between educational and income divides. 4. A large literature in economics and political science has documented the existence of unequal political representation and the distortion of politicians' and legislators' beliefs toward their most privileged constituencies: see Adams and Ezrow (2009); Gilens (2012); Bonica et al. (2013); Gilens and Page (2014); Kuhner (2014); Bartels (2017); Bertrand et al. (2020); Cagé (2020); Pereira (2021). We should stress, however, that the limitations of available information on party manifestos constrain our ability to carry out a causal analysis or fully test the hypotheses behind the empirical regularities we uncover. In particular, the sociocultural axis puts together many different items that may involve various forms of economic conflict over the consequences of environmental, migration, or education policies. The manifesto data do not provide information on the actual policies implemented by governing coalitions either. For instance, social democratic and affiliated parties may continue emphasizing redistributive policies just as they used to in the past, but their credibility in effectively pursuing these policies may have declined since then. Another complementary interpretation of our findings is that left-wing parties have gradually developed a more elitist approach to education policy, in the sense that they have increasingly been viewed by less well-off voters as primarily defending the winners of the higher education competition. 5 Unfortunately, the data at our disposal make it difficult to provide a direct test for these various hypotheses. The fact that turnout has fallen sharply among the bottom 50% least educated and poorest voters in a number of countries, but not among the top 50%, could be interpreted as a sign that socially disadvantaged voters have felt left aside by the rise of "multi-elite" party systems. 6 We investigate to what extent shifts in the composition of education groups in terms of gender, age, or other socioeconomic 5. This risk was identified as early as in 1958 by Michael Young in his famous dystopia about "the rise of the meritocracy." Young expresses doubts about the ability of the British Labour Party (of which he was a member) to keep the support of lower-educated classes in case the party fails to combat what he described as the rise of "meritocratic ideology" (a strong view held by higher education achievers about their own merit, which Young identified as a major risk for future social cohesion). For a simple theoretical model along these lines, see Piketty (2018, section 5). It is based on a two-dimensional extension of Piketty (1995)'s model about learning the role of effort and a distinction between education-related effort and business-related effort. The model can account for the simultaneous existence of Brahmin left voters (i.e., dynasties believing strongly in the role of educationrelated effort) and Merchant right voters (i.e., dynasties believing strongly in the role of business-related effort).
6. See Piketty (2018), figures A1 and A2. Turnout rates among bottom 50% voters have always been relatively low in the United States (at least after World War II). To some extent the British and French pattern has moved toward the U.S. pattern since the 1970s and 1980s. Unfortunately, the surveys at our disposal do not allow us to consistently analyze the evolution of turnout in our sample of 21 countries, so we do not push our analysis of turnout any further. variables could account for the reversal of the education cleavage. To do this, we compare the education gradient before and after controlling for all available covariates in our database. We carry a Kitagawa-Oaxaca-Blinder decomposition of the education gradient, which allows us to formally estimate what fraction of the reversal can be accounted for by structural changes in educational achievement. Both methods yield identical results: compositional effects can only predict 16%-17% of the transformation of educational divides observed since the 1950s.
We do find some heterogeneity in the reversal when further decomposing voters into subgroups by different socioeconomic characteristics. Generational dynamics appear to have mattered tremendously in generating the reversal of the education cleavage: while older lower-educated voters continue to vote "along class lines" and thus support the left, social democratic and green parties have attracted a growing share of the higher-educated electorate among the youth. The reversal in the educational divide has also been highest among nonreligious voters and among men, although it has happened in other subgroups, too. Overall, the disconnection of income and education cleavages has been a relatively independent and widespread phenomenon, in the sense that it cannot be accounted for by other socioeconomic variables and is not linked to any particular subgroup of voters.
Finally, we exploit the other variables in our data set to study cleavages related to age, geography, religion, gender, and other socioeconomic variables. The main conclusion is that there has been no major realignment of voters along these other dimensions comparable to the one observed in the case of education. Younger voters are more likely to vote for social democratic and affiliated parties, but this was already the case by a comparable magnitude in the 1950s. Similarly, rural-urban and religious cleavages have remained stable or have decreased in most countries in our data set: rural areas and religious voters continue to be supportive of conservative parties, as they used to be in the past. The major exception is gender, the only variable other than education for which we find a clear reversal of electoral divides: in nearly all countries, women used to be more conservative than men and have gradually become more likely to vote for left-wing parties.
This article directly relates to the growing literature on the sources of political change and the rise of "populism" in Western democracies. Recent studies have emphasized the role of various economic and sociocultural factors, including globalization and Downloaded from https://academic.oup.com/qje/article/137/1/1/6383014 by guest on 10 March 2022 trade exposure (Malgouyres 2017;Colantone andStanig 2018a, 2018b;Autor et al. 2020), economic insecurity and unemployment (Funke, Schularick, and Trebesch 2016;Algan et al. 2017;Becker, Fetzer, and Novy 2017;Becker and Fetzer 2018;Fetzer 2019;Liberini et al. 2019;Guiso et al. 2020;Dehdari forthcoming), immigration (Becker and Fetzer 2016;Halla, Wagner, and Zweimüller 2017;Dustmann, Vasiljeva, and Damm 2019;Tabellini 2020), and cultural and moral conflicts (Norris and Inglehart 2019;Enke 2020;Bonomi, Gennaioli, and Tabellini 2021). We contribute to this body of evidence by adopting a broader, long-run historical perspective on the evolution of political cleavages since the end of World War II. We find little evidence that the shifts in electoral divides we observe were driven by single, major events such as the end of the Cold War, the increasing salience of immigration since the 2000s, trade shocks, or the 2007-2008 crisis. What seems to have happened instead is a very progressive, continuous reversal of educational divides, which unfolded decades before any of these events took place and has carried on uninterruptedly until today.
We also contribute to the literature on multidimensional political competition and its effect on redistribution and inequality. A key result from this literature is that political support for redistribution should be inversely proportional to the strength of other political cleavages crosscutting class divides (Roemer 1998;Alesina, Baqir, and Easterly 1999;Roemer, Lee, and Van der Straeten 2007;Bonomi, Gennaioli, and Tabellini 2021). The divergence of the effects of income and education on the vote documented in this article, two highly correlated measures of inequality, could contribute to explaining why the rise of economic disparities in the past decades has not been met by greater redistribution or renewed class conflicts.
Finally, this article relates to the large political science literature on the determinants of the vote in comparative and historical perspective. Numerous studies have highlighted that Western democracies have undergone a process of growing polarization over a new "sociocultural," "universalistic-particularistic," or "green/alternative/libertarian versus traditional/authoritarian/ nationalist" dimensions of political conflict in the past decades (see Inglehart 1977;Kitschelt 1994;Hooghe, Marks, and Wilson 2002;Kriesi et al. 2008;Bornschier 2010a;Evans and De Graaf 2013;Dalton 2018;Norris and Inglehart 2019). There is also extensive evidence that education has been playing a major role in restructuring electoral behaviors and collective beliefs along this new dimension in recent decades (see Duch and Taylor 1993;Van der Waal, Achterberg, and Houtman 2007;Stubager 2008Stubager , 2010Stubager , 2013Bornschier 2010b;Dolezal 2010;Wille and Bovens 2012;Rydgren 2013Rydgren , 2018Kitschelt and Rehm 2019;Langsaether and Stubager 2019;Ford and Jennings 2020). We contribute to this literature by gathering the largest data set ever built on the socioeconomic determinants of the vote in Western democracies; 7 by focusing explicitly on the distinction between income and education, two variables whose effects are rarely studied jointly in comparative studies; and by directly matching this data set with historical data on party ideology to document the dynamic links between political supply and demand. 8 In doing so, we confirm many of the findings of the existing literature, but we also provide new insights into the transformation of political cleavages in Western democracies. In particular, for the first time we gather cross-country, long-run historical evidence of a gradual dissociation of the effects of education and income on the vote. This dissociation appears to have started as early as the 1950s and to have unfolded uninterruptedly since then, and it can be related to the growing salience of a large and complex set of policy issues, including the environment, migration, gender, education, and merit, which divide voters along educational but not income lines.
Section II presents the new data set exploited in this article. Section III documents the divergence of the income and education effects and discusses the role of green and anti-immigration parties in explaining the reversal of the education cleavage. Section IV matches our survey data set with manifesto data to study the link between this transformation and the emergence of a new axis of political conflict. Section V explores alternative explanations and heterogeneity in the reversal of the education cleavage and analyzes the evolution of other determinants of electoral behaviors. Section VI concludes. 7. Our work directly draws on previous data collection and harmonization efforts. See in particular Franklin et al. (1992), Thomassen (2005), Elff (2007), Evans and De Graaf (2013), Bosancianu (2017), Schmitt (2021), and the collections of postelectoral surveys compiled by the Comparative Study of Electoral Systems (http://cses.org) and the Comparative National Elections Project (https://u.osu.edu/cnep/). 8. In matching survey and manifesto data, we follow recent political science studies seeking to understand how political supply influences class and religious divides. See in particular Elff (2009) ;Need (2011, 2012); Tilley (2012, 2017); Evans and de Graaf (2013) ;Jansen, Evans, and de Graaf (2013); and Rennwald and Evans (2014).

II.A. A New Data Set on Political Cleavages in Western
Democracies , 1948-2020 The data set we exploit in this article consists of a collection of electoral surveys conducted between 1948 and 2020 in Western democracies. These surveys have one main point in common: they contain information on the electoral behaviors of a sample of voters in the last (or forthcoming) election, along with data on their main sociodemographic characteristics such as income, education, or age. While they suffer from limitations typical to surveys (such as small sample sizes), they provide an invaluable source for studying the long-run evolution of political preferences in contemporary democracies.
1. Universe. Our area of study encompasses 21 countries commonly referred to as Western democracies, for which we can cover a total of about 300 national elections (see Table I). These include 17 Western European countries, the United States, Canada, Australia, and New Zealand. For seven countries in our data set (France, Germany, Italy, Norway, Sweden, the United Kingdom, and the United States), available surveys allow us to go back as early as the 1950s. The majority of remaining countries have data going back to the 1960s or the early 1970s, with the exception of Spain and Portugal, which did not hold democratic elections between the 1940s and the late 1970s.
Our focus is on national (general or presidential) elections, which determine the composition of government and the head of the state. In the majority of Western democracies, these elections have been held on a regular basis every four or five years since at least the end of World War II. Depending on their frequency and the availability of electoral surveys, we are able to cover political attitudes in 9-21 of these elections in each country.
2. Data Sources. Our primary data source consists in socalled national election studies, most of which have been conducted by a consortium of academic organizations (see Table I). The majority of these surveys are postelectoral surveys: they are fielded shortly after the corresponding national election has been held, with sample sizes generally varying between 2,000 and 4,000 respondents, and they collect detailed information on voting behaviors and the sociodemographic characteristics of voters. In all Western democracies except Austria, Ireland, and Luxembourg we have been able to get access to such high-quality data sources. For these three countries, we rely on more general political attitudes surveys, which were not specifically conducted in the context of a election but did ask respondents to report their previous voting behaviors: the Eurobarometers, the European Social Survey, and the European Election Studies. Furthermore, in a few countries such as Australia or Belgium, where national election studies were not conducted before the 1970s or 1980s, we complement them with other political attitudes surveys conducted in earlier decades. Although these sources do not allow us to accurately track election-to-election changes, they are sufficient to grasp long-run changes in party affiliations, which is the objective of this study. 9 3. Harmonization. Starting from raw data files, we extract in each survey all sociodemographic characteristics that are sufficiently common and well measured to be comparable across countries and over time. Based on these criteria, we were able to build a harmonized data set covering the following variables: income, education, age, gender, religious affiliation, church attendance, race or ethnicity (for a restricted number of countries), rural-urban location, region of residence, employment status, marital status, union membership, sector of employment, homeownership, selfperceived social class, and (in recent years) country of birth.
Income and education, the two variables that form the core part of our analysis in Section III, deserve special attention. Indeed, one reason income and education variables are not often studied jointly in large-scale comparative studies on electoral behaviors is that they tend to be difficult to harmonize. Education systems and educational attainments vary significantly across countries and over time, and they are not always perfectly comparable across surveys. The same limitations apply to income, which is only collected in discrete brackets in the majority of the sources used in this article.
We address this shortcoming by normalizing these two variables and focusing on specific education and income deciles. Online Appendix A introduces the method we use to move from discrete categories (education levels or income brackets) to deciles. 9. A complete list of all data sources used by country can be found in Online Appendix Table A1.
In broad strokes, our approach consists in allocating individuals to the potentially multiple income or education deciles to which they belong, in such a way that average decile-level vote shares are computed assuming a constant vote share in each education-or income-year cell. This is a conservative assumption, as vote shares for specific parties are likely to also vary in education groups or income brackets. The levels and changes in education and income cleavages documented herein should thus be considered as lower bounds of the true effects of education and income on the vote.
Last, to make surveys more representative of election outcomes, we systematically reweight respondents' answers to match official election results. Given that postelectoral surveys capture relatively well variations in support for the different parties, this correction leaves our results unchanged in the majority of cases.

II.B. Party Classification
Our objective is to compare the long-run evolution of electoral cleavages in Western democracies. This requires grouping political parties in such a way that the size of the coalitions considered and their historical affiliations are as comparable and meaningful as possible. To do so, we make a distinction between two large groups of parties in our main specification (see the coalitions delineated by dashed lines in Figure IV). 10 On one side of the political spectrum are social democratic, socialist, communist, and green parties, often classified as leftwing and that we also refer to as social democratic and affiliated parties in what follows. These include the Democratic Party in the United States; labor parties in countries such as the United Kingdom, Australia, or Norway; as well as various parties affiliated to socialist and social democratic traditions in Western European countries. It also includes environmental parties in their various forms, together with several new left-wing parties that emerged after the 2008 crisis (such as Podemos in Spain, Die Linke in Germany, or La France Insoumise in France). Tables A2 and A3 for more information on the classification of the main parties in each country. Parties not classified in either of these groups mainly correspond to independent candidates and regional parties (such as the Bloc Québécois in Canada or the Scottish National Party). These parties or candidates have received about 7% of votes since 1945, with no clear trend (see Figure IV). On the other side are conservative, Christian democratic, and anti-immigration parties, often classified as right-wing and that we also refer to as conservative and affiliated parties. These include the Republican Party in the United States and other conservative parties such as those of the United Kingdom, Norway, and Spain; Christian democratic parties, which are common in Western European multiparty systems such as those of Austria, Belgium, and Switzerland; and anti-immigration parties such as the French Rassemblement National or the Danish People's Party. We also include parties commonly classified as liberal or social-liberal in this group, such as the Liberal Democrats in Britain, the Free Democratic Party in Germany, and the Liberal Party in Norway, but our results are robust to not doing so. 11 This binary classification has one major advantage: it allows us to directly compare electoral divides in two-party systems, such as the United Kingdom or the United States, to those observed in highly fragmented party systems such as France or the Netherlands. Aggregating parties into two large groups of comparable size in each country is thus useful to get a first perspective on the long-run evolution of political cleavages that is consistent both over time and across countries. These groups also correspond in many cases to the coalitions of parties that have effectively built political majorities, whether in coalition governments or through direct parliamentary support.

See Online Appendix
To make sure that this distinction between "left" and "right" is meaningful when it comes to differentiating parties and voters, we contrast two indicators for all parties: the average self-reported left-right position of voters supporting each of these parties, and the score of each of these parties on the left-right ideological index available from the Comparative Manifesto Project (CMP) database. The first of these indicators is available in most postelectoral surveys used in this article, which have directly asked respondents to position themselves on a 0 (left) to 10 (right) scale.
11. The exceptions are Austria, Canada, Denmark, and the Netherlands, for which we classify as "left-wing" parties generally considered to be liberal (NEOS in Austria, the Liberal Party in Canada, the Social Liberal Party in Denmark, and D66 in the Netherlands). This choice is motivated by our objective to compare coalitions of significant and comparable size across countries. Liberal parties have received about 10% of the vote in Western democracies since 1945 (see Figure IV), with no clear trend, and have consistently been supported by both high-income and higher-educated voters (see Online Appendix Figures A26 and A28). Our results are thus robust to excluding them or not from the analysis.
The second is a measure of parties' left-right positions that theoretically ranges from −100 (right) to 100 (left). It was first computed from manifesto data and validated by factor analysis by Budge and Laver (1992), and it has been widely used in comparative political science research since then (e.g., Evans and De Graaf 2013).
We find that our categorization of political parties into two groups is very consistent with these two indicators. Every single party that we have classified as "social democratic and affiliated" is supported by voters who declare being more left-wing than the average voter and is more left-wing than the average party on the CMP left-right ideological index. 12 This is true for social democratic and socialist parties, but also for green parties, which are all ranked as left-wing in survey and manifesto data. The same holds in the case of conservative, Christian democratic, and antiimmigration parties, which are nearly all identified as more rightwing than the average party or voter. Moreover, the two indicators of parties' positions on a left-right scale are also consistent with one another (the correlation between the variables is 0.82). We are thus confident that our classification is meaningful in terms of both parties' programmatic supply and voters' own perceptions of the political space.
That being said, we are not claiming that these two groups are ideologically or programmatically homogeneous in any way, neither internally nor over time. Our objective is, on the contrary, to document how such large families or parties have aggregated diverse and changing coalitions of voters in the past decades. In Section III, we thus consider in greater detail how specific subfamilies of parties (in particular, green and anti-immigration movements) have contributed to reshaping electoral divides in countries with multiparty systems.

II.C. Empirical Strategy
In the rest of the article, we present results from simple linear probability models of the form: (1) 12. See Online Appendix Figures B16 and B17. The one single exception here is Fianna Fáil in Ireland, which we still choose to classify with left-wing parties to study a coalition of sufficient size (if we were to exclude it, the total vote share of the "left" would fall below 30% throughout the period considered). where y ict is a binary outcome variable of interest (e.g., voting for left-wing parties) for individual i in country c in election t, x ict is a binary explanatory variable of interest (e.g., belonging to top 10% educated voters), and C ict is a vector of controls.
In the absence of controls, the coefficient β simply equals the difference between the share of top 10% educated voters voting for left-wing parties and the share of other voters (bottom 90% educated voters) voting for left-wing parties: With controls, the interpretation is also straightforward: all things being equal, belonging to the top 10% of educated voters increases one's propensity to vote for left-wing parties by β percentage points. All control variables in our data set are specified as dummy variables, so the model is fully saturated and can be estimated by OLS using heteroskedasticity-robust standard errors.

III. THE DISCONNECTION OF EDUCATION AND INCOME CLEAVAGES IN WESTERN DEMOCRACIES
This section presents our main results on the evolution of electoral divides related to income and education. Section III.A documents the reversal of the education cleavage and the stability of income divides. Section III.B studies how the fragmentation of party systems and the rise of green and anti-immigration parties has contributed to this transformation.

III.A. The Divergence of Income and Education
To document the evolution of the influences of income and education on the vote, we start by relying on a simple indicator: the difference between the share of the 10% most educated voters and the share of the 90% least educated voters voting for social democratic, socialist, communist, and green parties (that is, β in equation (1)). We use the same indicator for income, defined as the difference between the share of richest 10% voters and the share of poorest 90% voters voting for social democratic and affiliated parties. These two indicators have the advantage of measuring FIGURE I

The Disconnection of Income and Education Cleavages in Western Democracies
In the 1960s, higher-educated and high-income voters were less likely to vote for left-wing (social democratic/socialist/communist/green/other left-wing) parties than were lower-educated and low-income voters by more than 10 percentage points. The left vote has gradually become associated with higher education voters, giving rising to a complete divergence of the effects of income and education on the vote. Figures correspond to five-year averages for Australia, Britain, Canada, Denmark, France, Germany, Italy, the Netherlands, Norway, Sweden, Switzerland, and the United States. Estimates control for income/education, age, gender, religion, church attendance, rural/urban, region, race/ethnicity, employment status, and marital status (in country-years for which these variables are available). Data from World Political Cleavages and Inequality Database. the evolution of the voting behaviors of two groups of equal size, which makes the estimates more comparable. 13 Figure I depicts the average quinquennial evolution of these indicators, after controls, in the 12 Western democracies for which data is available since the 1960s. 14 As shown in the upper line, highest-educated voters were less likely to vote for social democratic parties than were lowest-educated voters by 15 percentage points in the 1960s. This gap has shifted very gradually from being negative to becoming positive, from −10 in the 1970s to −5 in the 1980s, 0 in the 1990s, +5 in the 2000s, and finally +10 in 13. As discussed in Section II.A, deciles of education are computed using all educational categories available in surveys, which implies that the composition of "top 10% educated voters" changes over time. At the beginning of period, this category is mainly composed of university graduates and voters with secondary education; in the 2010s, it gives more weight to individuals with master's or doctorates. See Online Appendix A for more details.
14. The corresponding regression coefficients by country and decade are displayed in Online Appendix Tables D1 and D2.
Downloaded from https://academic.oup.com/qje/article/137/1/1/6383014 by guest on 10 March 2022 2016-2020. Higher-educated voters have thus moved from being significantly more right-wing than lower-educated voters to significantly more left-wing, leading to a complete reversal in the educational divide. The evolution has been dramatically different in the case of income. The bottom line shows that top-income voters have always been less likely to vote for social democratic and affiliated parties than low-income voters. This gap has decreased from −15 in the 1960s to about −10 in the past decade, but it remains negative. High-income voters have thus remained closer to conservative parties than low-income voters over the past 50 years.
Combining these two evolutions, a striking long-run transformation in the structure of political cleavages emerges. In the early postwar decades, the party systems of Western democracies were "class-based," in the sense that social democratic and affiliated parties represented both the low-education and the lowincome electorate, whereas conservative and affiliated parties represented both high-education and high-income voters. These party systems have gradually evolved toward what we propose to call "multiconflictual" or "multi-elite" party systems: higher-educated voters now vote for the left, while high-income voters still vote for the right.
Note that the two indicators in the figure control for all available variables at the micro level (education/income, age, gender, religion, church attendance, rural/urban location, region, race/ethnicity, employment status, and marital status). The evolution of these indicators without controls displays a larger decline in the influence of income on the vote, from nearly −20 in the 1960s to about −5 in 2016-2020. The main reason is that highereducated voters have on average higher incomes, so the reversal of the educational divide has mechanically led to a reduction in the difference between top-income and low-income voters. Nonetheless, what is important for our analysis is that the transition observed is robust to the inclusion or exclusion of controls. 15 The divergence of divides related to income and education is common to nearly all Western democracies, but it has happened at different speeds and with different intensities. Figure II shows that the support of higher-educated voters for social democratic parties was lowest in Norway, Sweden, and 15. See Online Appendix Figure A1. We come back to the influence of other covariates in generating the evolution of the education cleavage in Section V.

FIGURE II
The Reversal of Educational Divides in Western Democracies The figure represents the difference between the share of higher-educated (top 10%) and lower-educated (bottom 90%) voters voting for social democratic/socialist/ communist/green/other left-wing parties in English-speaking and Northern European countries (Panel A) and Continental and Southern European countries (Panel B). In nearly all countries, higher-educated voters used to be significantly more likely to vote for conservative parties and have gradually become more likely to vote for these parties. Estimates control for income, age, gender, religion, church attendance, rural/urban, region, race/ethnicity, employment status, and marital status (in country-years for which these variables are available). Data from World Political Cleavages and Inequality Database.
Finland between the 1950s and 1970s, three democracies well known for having stronger historical class-based party systems than most Western democracies. The reversal of the education cleavage has not yet been fully completed in these countries, as social democratic parties have managed to keep a nonnegligible This delay is also common to recent democracies, such as Spain or Portugal, or late industrialized countries such as Ireland, where left-wing parties continue to be more class-based. Portugal and to a lesser extent Ireland represent two major exceptions in our data set, where we do not observe a clear tendency toward a reversal of the educational divide. Among several factors, this unique trajectory can be explained by the polarization of mainstream parties and the success of new left-wing parties after the onset of the 2008 financial crisis (Bauluz et al. 2021). In contrast, the gap in left votes between higher-educated voters and lower-educated voters is today highest in countries such as the United States, Switzerland, and the Netherlands, largely due to the particular salience of identity-based concerns and the strength of anti-immigration and green movements in the latter two countries (Durrer de la Sota, Gethin, and Martínez-Toledano 2021). Figure III shows that top-income voters have also remained more likely than low-income voters to vote for conservative and affiliated parties in nearly all Western democracies, but with important variations. The influence of income on the vote was largest in northern European countries, Britain, Australia, and New Zealand in the 1950s and 1960s, consistent with their histories of early industrialization and class polarization. It has declined in these countries since then, although income continues to be negatively associated with support for the left.
Meanwhile, low-income voters have less decisively supported left-wing parties in countries with weak historical class cleavages and crosscutting religious (Italy) or ethnolinguistic (Canada) cleavages (Bauluz et al. 2021;Gethin 2021). Despite these variations, the tendency of high-income voters to support the right in contemporary Western democracies has proved remarkably resilient over time, pointing to the persistence of conflicts over economic issues and redistributive policy. The only country where a flattening of the income effect could well be underway is the United States (as well as Italy, due to the recent success of the Five Star Movement among the low-income electorate), where in 2016 and 2020, for the first time since World War II, top 10% earners became not significantly less likely to vote for the Democratic Party.
Our findings on the reversal of educational divides and the stability of the income effect are extremely robust to alternative specifications. The pattern observed is virtually identical whether one considers the top 50% of education and income voters, other

FIGURE III
The Stability/Decline of Income Divides in Western Democracies The figure represents the difference between the share of high-income (top 10%) and low-income (bottom 90%) voters voting for social democratic/socialist/ communist/green/other left-wing parties in English-speaking and Northern European countries (Panel A) and Continental and Southern European countries (Panel B). In all countries, top-income voters have remained significantly less likely to vote for these parties than low-income voters. Estimates control for education, age, gender, religion, church attendance, rural/urban, region, race/ethnicity, employment status, and marital status (in country-years for which these variables are available). Data from World Political Cleavages and Inequality Database. discrete categories such as primary-educated voters or university graduates, or continuous measures of education and income, before and after controls. 16  only in relative terms: between 1948-1960 and 2016-2020, for instance, the share of least educated 50% voters voting for social democratic and affiliated parties declined from about 50% to 40%, whereas it rose linearly from 25% to almost 50% among the top 10% educated (see Online Appendix Figure A29). We also find that our results hold when considering a continuous measure of leftright voting derived from the CMP database instead of a binary dependent variable (see Online Appendix Tables D5-D8). Finally, we report in the Online Appendix full regression tables on the determinants of the vote for social democratic and affiliated parties by country, as well as simple descriptive statistics on support for these parties by education level and income group in each country. 17 With the exception of the few cases highlighted above, we find a complete reversal of the education effect and a stability of the income effect in nearly all countries, regardless of the indicator considered to measure the influence of these two variables. 18

III.B. The Fragmentation of Political Cleavage Structures
The emergence of multiparty systems has come together with a significant reshuffling of political forces in most Western democracies. As shown in Figure IV, traditional socialist and social democratic parties have seen their average vote share across Western democracies decline from about 40% to 34% since the end of World War II, and that received by Christian democratic and conservative parties has decreased from 38% to 30%. Communist parties, who used to gather 7% of the vote in the 1940s, have available in each survey. If 25% of voters are primary educated, 50% are secondary educated, and 25% are tertiary educated, for instance, then voters belonging to each of these categories are attributed quantile values of 0, 0.25, and 0.75, respectively. 18. In some cases, the effect of income is nonlinear, especially at the beginning of the period: support for left-wing parties is higher among middle-income groups than at the bottom of the distribution. This is mainly because farmers and the selfemployed, many of whom have low incomes, have always been substantially more likely to vote for conservative parties. However, income remains an only imperfect and partial measure of economic resources. In particular, we find in the case of France (the only country with high-quality wealth data) that the effect of wealth on support for the left is much larger and linear, and has remained more stable in the past decades (see Online Appendix Figures EC1 and EC2).

FIGURE IV
The Transformation of Western Party Systems, 1945-2020 The figure represents the average share of votes received by selected families of political parties in Western democracies between the 1940s and the 2010s. Communist parties saw their average scores collapse from 7% to less than 0.5%, while green and anti-immigration parties rose until reaching average vote shares of 8% and 11%, respectively. Decennial averages over all Western democracies except Spain and Portugal (no democratic elections before 1970s) and the United States and the United Kingdom (two-party systems). The dashed lines delimit the categorization of parties considered in the main specification (social democrats and affiliated, conservatives and affiliated, and other parties). Data from official election results. almost completely disappeared from the political scene. Although immigration issues were already present in political debates in many Western democracies, anti-immigration parties started to grow in the late 1970s and have seen their support increase uninterruptedly since then, reaching on average 11% of votes in the past decade. Green parties made their entry in the political landscape in the 1970s and 1980s and have also progressed steadily, reaching on average 8% of votes in the past decade. Support for social-liberal and liberal parties has remained more stable, even though there are important variations across countries. Figure V displays the evolution of our previous education (Panel A) and income (Panel B) indicators, decomposed for each family of parties from 1948 to 2020. In the 1950s and 1960s, both top 10% educated voters and top 10% income voters were significantly less likely to vote for social democratic, socialist, communist, and other left-wing parties and more likely to vote for conservative, Christian democratic, and liberal parties than other voters. By 2016-2020, income continues to clearly distinguish The estimates are presented after controlling for income, age, gender, religion, church attendance, rural/urban, region, race/ethnicity, employment status, and marital status (in country-years for which these variables are available). Panel B: The estimates are presented after controlling for education, age, gender, religion, church attendance, rural/urban, region, race/ethnicity, employment status, and marital status (in country-years for which these variables are available). Data from World Political Cleavages and Inequality Database. these groups of parties, but their education gradient has completely reversed. Meanwhile, support for anti-immigration and green parties does not differ significantly across income groups (their income gradient is close to zero), but it does vary substantially across educational categories. This has been a constant fact since these parties started taking on a growing importance in the political space in the 1970s and 1980s. In 2016-2020, top 10% educated voters were more likely to vote for green parties by 5 percentage points and less likely to vote for anti-immigration parties by a comparable amount. In other words, the increasing support for green parties on the left and anti-immigration parties on the right has clearly contributed to the reversal of the education cleavage. This finding is in line with the large political science literature that has shown education to be an important determinant of support for green and anti-immigration parties in recent years (e.g., Dolezal 2010; Rydgren 2013, 2018; Abou-Chadi and Hix 2021).
We should stress, however, that the rise of new parties alone cannot explain the reversal of the education cleavage for at least two reasons. First, this reversal started several decades before most of these parties even existed: as Figure V shows, we can date it back to as early as the 1950s. Second, as shown in Figure V, there have been major transformations in the structure of the vote for traditional left-wing and right-wing parties, too, even in the most recent decades. One way to formally decompose the respective influences of traditional left-wing parties and green parties in generating the reversal of the education cleavage is to compare our main indicator of interest including and excluding green parties from the analysis. We find that the gradient has moved from −19.1 to +8.2 between 1948-1960 and 2016-2020 when including green parties, and from −19.1 to +4.3 when excluding them. In other words, the rise of green parties explains about 15% of the reversal observed during this period, and it explains about half of the positive link between education and support for the left in the most recent years. The same holds when it comes to the increase in support for anti-immigration parties in generating the reversal of the link between education and support for the right: it explains about 14% of the overall shift and 55% of the negative gradient in 2016-2020. 19 Figure VI provides another perspective on this transformation by representing the income and education gradients of these different families of parties in a two-dimensional space in 1961-1965 (Panel A) and 2016-2020 (Panel B). In the 1960s, the effects The figure represents the difference between the share of high-income (top 10%) and low-income (bottom 90%) voters voting for selected groups of parties on the y-axis, and the same difference between higher-educated (top 10%) and lowereducated (bottom 90%) voters on the x-axis. In the 1960s, social democratic, socialist, and communist parties were supported by low-income and lower-educated voters, while conservative, Christian, and liberal parties were supported by highincome and higher-educated voters. By 2016-2020, education most clearly distinguishes anti-immigration from green parties, while both income and education most clearly distinguishes conservative and Christian democratic parties from socialist, social democratic, and communist parties. Averages over all Western democracies. Estimates control for income/education, age, gender, religion, church attendance, rural/urban, region, race/ethnicity, employment status, and marital status (in country-years for which these variables are available). Data from World Political Cleavages and Inequality Database. of income and education on the vote were aligned: higher income and higher education were both associated with higher support for conservative and affiliated parties. By 2016-2020, these two variables now have opposite effects: higher income is associated with higher support for conservative parties, whereas higher education is associated with greater support for social democratic parties. Anti-immigration and green parties differ primarily in their tendency to attract voters belonging to different education groups (they are distant on the x-axis but not on the y-axis).
Figure VII further decomposes this two-dimensional structure of political conflict by country in the past decade, distinguishing between traditional right-wing and left-wing parties in Panel A and between anti-immigration and green parties in Panel B. 20 The two-dimensional split of the electorate can be seen in nearly all countries in our data set: social democratic and other left-wing parties systematically make better relative scores among low-income voters, conservative and other right-wing parties among high-income voters, anti-immigration parties among lower-educated voters, and green parties among higher-educated voters. 21 Despite these commonalities, there are large differences across countries in these indicators. In particular, while nearly all green parties make better scores among higher-educated voters than among the lower educated, they differ in their tendency to attract low-or high-income voters. Similarly, anti-immigration parties have attracted a particularly high share of the lowereducated vote in several Western democracies in the past decade, but we also observe variations in the income profile of far-right voting. These variations are likely to reflect cross-country differences in political fragmentation and voting systems, which create different incentives for parties of the traditional left and the 20. The corresponding regression coefficients by country and decade, after controls, are displayed in Online Appendix Tables D3 and D4. 21. In Italy and New Zealand, lower-educated voters are not significantly more or less likely to vote for anti-immigration parties. In Italy, this is driven by the fact that support for Fratelli d'Italia (which we classify as an anti-immigration party alongside the Lega) was particularly concentrated among higher-educated voters in the 2018 election. In New Zealand, the only significant anti-immigration party, New Zealand First, receives support mainly from the Māori minority and is often considered to be a centrist party, which may explain why its position on the income-education quadrant differs from that of other anti-immigration parties (Gethin 2021).

FIGURE VII
Decomposing Income and Education Cleavages The figure represents the difference between the share of high-income (top 10%) and low-income (bottom 90%) voters voting for selected groups of parties on the y-axis, and the same difference between higher-educated (top 10%) and lowereducated (bottom 90%) voters on the x-axis, in the last election available (between 2014 and 2020). Estimates control for income/education, age, gender, religion, church attendance, rural/urban, region, race/ethnicity, employment status, and marital status (in country-years for which these variables are available). Data from World Political Cleavages and Inequality Database.
traditional right to adapt their policy proposals in the face of growing electoral competition from new political movements. To better understand these dynamics and the role of political supply in shaping education and income divides, we now turn to manifesto data.

IV. THE ORIGINS OF THE TRANSFORMATION OF POLITICAL CLEAVAGES: EVIDENCE FROM MANIFESTO DATA
This section investigates the relationship between the divergence of income and education cleavages and ideological polarization by matching our survey data set with manifesto data. Section IV.A introduces the CMP data and the indicators we consider. Section IV.B presents our results on the link between political supply and demand.

IV.A. Manifesto Project Data and Methodology
1. Manifesto Data. To make a first step toward understanding the mechanisms underlying the transformation documented in Section III, we match our survey data set with the CMP (Volkens et al. 2020), a hand-coded historical database on the programmatic supply of political parties. The CMP is the result of a collective effort to collect and code the manifestos published by parties just before general elections. Each manifesto is first divided into quasi-sentences conveying a specific claim or policy proposal. These quasi-sentences are assigned to broad ideological or policy categories using a common coding scheme. The resulting data set presents itself in the form of items (such as "social justice" or "law and order"), with scores corresponding to the share of quasi-sentences dedicated to a specific issue in a party's manifesto. The CMP is the largest available database on political programs in contemporary democracies at the time of writing, and the only one covering nearly all elections held in our 21 countries of interest since the end of World War II.

Combination of Manifesto and Survey Data.
We proceed by matching one by one every party reported in both the CMP and our data set. This was possible for a total of 459 parties, allowing us to cover over 90% of votes cast in nearly all elections in the survey data. The remaining correspond either to independent candidates, or to small parties for which data was not available in the CMP. To the best of our knowledge, this represents the most comprehensive mapping between political supply and demand ever built in comparative research.
3. Indicators of Interest. Following the political science literature, we consider two main indicators of political supply proposed by Bakker and Hobolt (2013). The indicators correspond to Downloaded from https://academic.oup.com/qje/article/137/1/1/6383014 by guest on 10 March 2022 parties' positions on two axes of political conflict: an "economicdistributive" axis representing divides over economic policy and inequality, and a "sociocultural" axis mapping conflicts over issues such as law and order, the environment, multiculturalism, or immigration. 22 The economic-distributive indicator is equal to the difference between the percentage of "pro-free-market" statements and "pro-redistribution" statements in a given party's manifesto. Pro-redistribution emphases include proposals to expand social services, nationalize industries, or enhance social justice. Meanwhile, pro-free-market statements encompass references to the limitation of social services, economic incentives, and free enterprise.
Conversely, the sociocultural indicator is defined as the difference between the percentage of "progressive" emphases and "conservative" emphases. Conservative emphases include categories such as political authority, positive evaluations of traditional morality, or negative attitudes toward multiculturalism. Progressive emphases cover issues related to environmentalism, the protection of minority groups, or favorable mentions of multiculturalism.
Given that manifesto items sum by definition to 100%, both indicators theoretically range from −1 to 1, with 1 representing a case of a party exclusively emphasizing pro-freemarket/conservative values, and −1 that of a party exclusively emphasizing pro-redistribution/progressive values. Although these measures of political ideology remain broad and are not exempt from measurement error, they represent the best data at our disposal to study the link between political supply and demand in the long run.
We also stress that by operating this distinction between economic and sociocultural dimensions of political conflict, we are not suggesting that sociocultural divides are purely conflicts over identity or morality that would be fully exempt from material concerns. Immigration, environmental, and cultural policies can have strong distributional implications, for instance by disproportionally affecting low-skilled workers or by mostly benefiting residents of large cities, who tend to concentrate a larger share of the higher-educated electorate. In that respect, the emergence of a secondary dimension of political conflict linked to education should also be understood as incorporating new forms of socioeconomic cleavages.

IV.B. The Evolution of Ideological Polarization
How has the structure of economic and sociocultural conflicts changed in Western democracies since the end of World War II, and to what extent can this account for the growing disconnection between the influences of income and education on the vote? Figure VIII provides a first answer to this question by displaying the evolution of the average economicdistributive and sociocultural scores of specific families of parties between 1945 and 2020. 23 Indices are normalized by the average score by decade so as to better highlight the dynamics of polarization.
Polarization on economic issues has remained remarkably stable in the past decades. The economic-distributive score of social democratic and socialist parties has remained 9-14 points below average, while that of conservative parties has fluctuated between +8 and +11. Green parties, which started gaining electoral significance at the beginning of the 1980s, have held economic positions that are comparable to that of traditional leftwing parties. Anti-immigration parties have moved closer to the average position of conservative parties, after a period of particularly marked emphasis on pro-free-market policies. This is consistent with qualitative accounts of the ideological transformation of far-right movements in Western Europe, from the Freedom Party of Austria (Durrer de la Sota,  to the French Rassemblement National (Piketty 2018) and the True Finns (Martínez-Toledano and Sodano 2021), which have shifted to defending redistributive economic policies in recent years.
Meanwhile, polarization on the sociocultural axis of political conflict has dramatically risen since the 1970s, after a brief period of convergence in the early postwar decades. This polarization has been driven by both old and new parties. Between 1970 and 2020, social democratic and socialist parties increasingly emphasized progressive issues, as their deviation from the mean sociocultural 23. The underlying figures are reported in Online Appendix Table B2. See Online Appendix Figures B2-B8 for a complete representation of the political space by decade.

FIGURE VIII
The Evolution of Ideological Polarization in Western Democracies, 1945Democracies, -2020 Panel A displays the average economic-distributive scores by decade for four families of parties across all Western democracies: social democratic, socialist, communist, and other left-wing parties; conservative, Christian democratic, and liberal parties; anti-immigration parties; and green parties. Negative values on the economic-distributive index correspond to greater proportions of pro-redistribution emphases relatively to pro-free-market emphases. Panel B displays the average sociocultural scores by decade for the four families of parties. Negative values on the sociocultural index correspond to greater proportions of progressive emphases relatively to conservative emphases. Indices are normalized by the average score by decade so as to better highlight the dynamics of polarization. Data from the Comparative Manifesto Project database. score declined linearly from −0.6 to −5.4, while conservative parties shifted to more conservative positions. Green parties have consistently emphasized progressive issues to a much greater extent than other parties since their emergence in the 1980s, with a stable score of about −25. Finally, anti-immigration parties have seen their score on the sociocultural axis surge, from +4 in the 1970s to +20 in the 2010s.
Beyond these two indicators of party ideology, we provide more detailed results on the structure of the manifestos of each party family in the Online Appendix (see Tables B3-B7). Two key results stand out from these disaggregated figures. First, the conservative turn of anti-immigration and other right-wing parties has been mainly driven by three items coded in the database: positive emphases of national way of life (including appeals to nationalism and patriotism), positive emphases of law and order (corresponding to favorable mentions of strict law enforcement and stricter actions against crime), and negative mentions of multiculturalism. 24 Meanwhile, green and other left-wing parties have dedicated a growing share of their manifestos to environmental issues and to positive emphases of an anti-growth economy (including calls for a more sustainable development path). Second, we find that left-wing and right-wing parties continue to differ on many issues on the economic-distributive dimension. In particular, green and other left-wing parties tend to put greater emphasis on welfare, equality, and social justice, whereas the manifestos of anti-immigration and other right-wing parties contain a larger share of sentences promoting a free-market economy and welfare state limitation.

IV.C. Ideological Polarization and the Transformation of Electoral Divides
The stability of economic-distributive conflicts and the rise of sociocultural divides resonates well with our finding on the stability of the income gradient and the reversal of the education cleavage. In particular, if the two phenomena are related, one might expect to observe that (i) parties with more progressive positions attract a relatively higher share of higher-educated voters, (ii) this relation should rise over time as sociocultural issues gained prominence, and (iii) countries that are more polarized on sociocultural issues should have higher education gradients, 24. Consistent with the idea that new ethnoreligious minorities perceive conservative and anti-immigration parties as particularly hostile to their integration, we find that immigrants and Muslim voters have been substantially more likely to vote for social democratic and affiliated parties than other voters in the past decade (see Online Appendix Figures CE1 and CE2). We also find deep and persistent divides between voters belonging to different racial or ethnic groups in countries with available data (see Online Appendix Tables E14, E20, and E21). The upper line plots the raw correlation between the education gradient (defined as the share of top 10% educated voters in the electorate of a given party) and the sociocultural index across all parties in the database. The bottom line plots the raw correlation between the income gradient (defined as the share of top 10% income voters in the electorate of a given party) and the economic-distributive index (inverted, so that higher values correspond to greater pro-redistribution emphases). The unit of observation is the political party. Error bars represent 95% confidence intervals. Data from the World Political Cleavages and Inequality Database and the Comparative Manifesto Project. thereby accounting for the cross-country variations documented in Section III. Figure IX provides descriptive evidence that the reversal of the education cleavage and the rise of a second dimension of political conflict are tightly associated. The upper line represents the correlation between the education gradient of a given party and the sociocultural index of this party by decade, computed across all parties available in the database. This correlation was close to 0 and not statistically significant in the 1960s. It has risen monotonically since then, from 0.1 in the 1970s to 0.3 in the 1990s and finally 0.46 in the past decade. Meanwhile, as represented in the bottom line, the correlation between the income gradient and the position of a given party on the economicdistributive axis has remained very stable and negative over the entire period. In other words, higher-educated voters have gradually converged in supporting parties with progressive positions, while high-income voters continue to vote for parties with pro-free-market positions just as much as they used to in the immediate postwar era. 25 We also investigate in greater detail how these correlations vary across all items available in the CMP database. 26 We find that the transformation documented above is visible in nearly all subcategories. In the 1960s-1970s, the education gradient was not significantly correlated to any of the items composing the sociocultural index. By 2010-2020, it has become strongly negatively correlated to positive emphases of law and order, national way of life, and traditional morality, and to negative mentions of multiculturalism. At the same time, it has become strongly positively correlated to positive emphases of culture, anti-growth economy, freedom and human rights, environmentalism, and multiculturalism. These results suggest that the emergence of a new sociocultural axis of political conflict cannot be narrowed down to a single topic of divergence: it involves conflicting visions and priorities over a complex and diverse set of issues. Figure X plots the cross-country relation between a simple measure of ideological polarization, defined as the standard deviation of the sociocultural index across all parties in a given election, and the education gradient in the past decade. The relation between the two indicators is strongly positive: countries in which parties compete more on sociocultural issues also display a greater propensity of higher-educated voters to support social democratic, socialist, green, and affiliated parties. In particular, we see that Portugal and Ireland, which were identified as exceptions showing no clear trend toward a reversal of the education cleavage, are the countries where sociocultural polarization is today the lowest. 27 Although the small number of countries makes it difficult to precisely identify the evolution of this relationship, we also find that it has grown over time, in line with our party-level analysis. 28 25. This transformation is robust to controlling for the composition of parties' electorates in terms of other variables, as well as to accounting for country, year, and election fixed effects (see Online Appendix Table B9).
26. See Online Appendix Table B10, which reports correlation coefficients between all items available in the CMP data set and our education and income indicators.
27. Notice that the indicator mechanically overestimates polarization in highly fragmented party systems such as that of Denmark, whereas it underestimates it in countries with fewer parties, such as the United Kingdom, New Zealand, and the United States. This may explain why these countries have lower levels of sociocultural polarization than one might expect.
28. See Online Appendix Figure B15, which reproduces Figure  Results combining data on political supply and demand therefore suggest that the emergence of a new sociocultural axis of political conflict is tightly linked to the reversal of the education cleavage in Western democracies. As parties have progressively come to compete on sociocultural issues, electoral behaviors have become increasingly clustered by education group. This relation holds at the country level, with the divergence between education and income being more pronounced in democracies where parties compete more fiercely on this new dimension of electoral divides.

V. ELECTORAL CHANGE IN WESTERN DEMOCRACIES: ALTERNATIVE EXPLANATIONS AND OTHER DIMENSIONS OF POLITICAL CONFLICT
This section studies alternative explanations and heterogeneity in the reversal of the education cleavage and analyzes other dimensions of political conflict. Section V.A investigates the extent to which the reversal of educational divides can be explained by changes in the composition of education groups. Section V.B explores heterogeneity in this reversal in terms of age, gender, religion, and other variables in our data set. Section V.C briefly discusses the evolution of other electoral cleavages in Western democracies, independently from education and income.

V.A. Can Compositional Changes Explain the Reversal of Educational Divides?
In previous sections, we studied the reversal of the education cleavage across all Western democracies, with little consideration for changes in the link between education and the other variables in our data set. Although we have shown that this reversal is robust to accounting for all available controls, the extent to which shifts in the composition of education groups could account for some of the transformation remains unclear. It is well known, for instance, that women have become both more educated (e.g., Vincent-Lancrin 2008;Parro 2012;Riphahn and Schwientek 2015) and more left-wing than men in the past decades (see Section V.C). The realignment of gender divides could thus have contributed to generating the move of higher-educated voters toward social democratic and affiliated parties. Similarly, the secularization of Western societies and the associated increase in the share of nonreligious voters, who tend to be more educated, could have facilitated the transformation of the education cleavage.
To investigate the role of these various factors, we conduct two complementary analyses: a comparison of the education gradient before and after controls, and a Kitagawa-Oaxaca-Blinder decomposition of the education cleavage. To derive meaningful comparisons, we restrict the analysis in this section to countries for which we have data since the 1960s and the richest comparable set of covariates (Australia, Denmark, Finland, France, the Netherlands, Norway, New Zealand, Sweden, the United Kingdom, and the United States).
We find that including control variables only marginally affects the overall change in the link between education and the vote of the past decades (see Online Appendix Table D9). More precisely, the top 10% educated voters were less likely to vote for social democratic and affiliated parties by 21.6 percentage points in the 1960s, while they were more likely to do so by 5.3 points in the 2010s. This represents an overall change in the education gradient of 26.9 percentage points. Adding controls does slightly affect the level of the coefficient, but it does not significantly Downloaded from https://academic.oup.com/qje/article/137/1/1/6383014 by guest on 10 March 2022 affect the trend: the education gap after controlling for all available covariates has moved from about −18.8 to 3.6, amounting to a shift of 22.4 percentage points. By this measure, changes in the composition of education groups can only account for about 16% of the transformation of educational divides.
Another, more formal way of evaluating what fraction of the reversal is due to changes in the composition of groups is to directly estimate a two-way Kitagawa-Oaxaca-Blinder decomposition of the education gradient (Kitagawa 1955;Blinder 1973;Oaxaca 1973). This allows us to decompose the marginal effect of education into two components: one that can be explained by group differences in predictors (that is, differences in the composition of education groups in terms of age, gender, etc.), and one that remains unexplained. As before, we find that other variables largely fail to account for the reversal of educational divides: the actual coefficient shifts from −22.5 to +10.4 between 1961-1965 and 2016-2020 (corresponding to a 32.9-point change), while the unexplained component increases from −19.6 to +7.6 (corresponding to a 27.2-point change). 29 This implies that these covariates can only predict 17% of the reversal observed over the period considered.

V.B. Heterogeneity in the Reversal of Educational Divides
Although compositional changes only explain between 16% and 17% of the reversal of the educational divide, we find some heterogeneity in the reversal when further decomposing voters into subgroups by different socioeconomic characteristics.
In particular, generational dynamics appear to have played a major role in the reversal of the education cleavage. Figure XI decomposes the evolution of the education gradient by cohort of voters born at different decades of the twentieth century. Highereducated voters have been more likely to vote for social democratic and affiliated parties than lower-educated voters in generations born after the 1940s, and the opposite is true among generations born before World War II. New generations have thus become increasingly divided along educational lines, suggesting that the education cleavage could continue rising in the future, as old generations voting along historical class lines gradually disappear The figure represents the difference between the share of higher-educated (top 10%) and lower-educated (bottom 90%) voters voting for social democratic/socialist/communist/green parties in specific cohorts of voters. Between the 1960s and the 1990s, lower-educated voters born in the early decades of the twentieth century remained significantly more likely to vote for these parties than were higher-educated voters born in the same period. In the past decade, on the contrary, young lower-educated voters were significantly less likely to vote for these parties than were young higher-educated voters. Figures correspond to 10 from the political landscape. The reversal of the education cleavage has, however, also taken place within recent cohorts, which points to the role of other factors potentially related to political supply or ideological change, as documented in Section IV.
We find some heterogeneity in the education gradient across other subgroups of voters (see Online Appendix Table D10). In the 2010s, the educational divide is higher among men than women, among nonreligious voters than religious voters, among public sector than private sector employees, and in rural areas than in urban areas. The reversal in the educational divide has also been highest among nonreligious voters and among men, although it has occurred in nearly all groups. Overall, this evidence reveals that although there exist interesting heterogeneities, the reversal of the educational divide has been a widespread phenomenon that is not restricted to a particular subgroup of voters.

V.C. The Evolution of Other Electoral Cleavages
We conclude by briefly discussing the evolution of other determinants of electoral behaviors. Our main finding is that there has been either a stability or a decline of their effect on vote choices. The major exception is gender, for which we find a significant reversal, comparable in magnitude to that of the education cleavage.
1. Generational Cleavages. Young voters have always been more likely to vote for left-wing parties than older cohorts in the majority of Western democracies. However, while there are fluctuations across countries and over time, we do not find any evidence that this cleavage has deepened in recent decades (see Online Appendix Figures CA1-CA4). We also document variations in the profile of the vote for anti-immigration parties by age across Western democracies: the share of votes received by these parties increases with age in Denmark, Italy, Norway, New Zealand, Switzerland, and Sweden, but it clearly decreases with age in Austria, Spain, Finland, and France (see Online Appendix Figures CA5-CA7). These findings call into question the strand of the political science literature that has argued that political change in Western democracies would have a major generational dimension, and that the emergence of populist authoritarian leaders in recent years would have partly represented a "backlash" against social progress among the older generations (see Inglehart 1977;Norris and Inglehart 2019). As shown in the previous section, educational divides within recent cohorts, rather than conflicts between generations, seem to represent a more important source of electoral realignment in contemporary democracies.
2. Rural-Urban Cleavages. We find that rural-urban divides have remained relatively stable in the past seven decades: rural areas continue to be more likely to vote for conservative and affiliated parties by 5 to 15 percentage points in most Western democracies, just as they used to in the 1950s and 1960s (see Online Appendix Figure CB1). Furthermore, the fragmentation of the political space in multiparty systems has been associated with a reshuffling of rural-urban divides within rather than across left-right blocs: support for green parties tends to be concentrated in cities today, just like other left-wing parties, while anti-immigration parties generally fare better in rural areas, as is the case of other conservative parties. The stability of the rural-urban cleavage thus rules out this dimension as the primary driver of electoral change since the end of World War II. 30 3. Religious Cleavages. Religious divides do not seem to have undergone any clear reversal in the past decades either. In all countries with available data, religious voters have always been much less likely to vote for social democratic and affiliated parties than nonreligious voters (see Online Appendix Figure CC1). This gap has slightly declined in most countries since the 1960s, but it remains decisively negative. Moreover, although green movements often disproportionately attract nonreligious individuals, this does not make them very different from other left-wing parties, which have always found greater support among secular voters. Support for anti-immigration parties appears to vary little across religious groups in most countries, so that their progression in recent decades has contributed to further weakening the religious cleavage (see Online Appendix Figure CC5 for green parties and CC6 for anti-immigration parties).

Gender Cleavages.
We also corroborate across all Western democracies a well-known fact (Inglehart and Norris 2000;Edlund and Pande 2002;Abendschön and Steinmetz 2014): women used to be more conservative than men and have gradually become more likely to vote for social democratic and affiliated parties (see Online Appendix Figure CD2). This transition, as in the case of the education cleavage, has been very gradual and is visible as early as the 1950s. In line with the existing literature, we find that much of the negative gradient of the early postwar decades can be explained by the fact that women used to be more religious than men (Blondel 1970;Goot and Reid 1984). In particular, this explains why the gender divide was exceptionally large in Italy in the 1950s, where religious cleavages were historically more pronounced than in most Western democracies. However, the reversal holds even after controlling for all available variables 30. The share of votes received by green and anti-immigration parties by rural-urban location is represented in Online Appendix Figures CB2 and CB3, respectively. Notice that a few Western democracies (in particular Australia, Belgium, Britain, and France) seem to have witnessed a significant transformation of center-periphery cleavages in recent years, as left-wing parties have concentrated a growing share of the vote in capital cities (see Online Appendix Figures  CB4-CB8).
Downloaded from https://academic.oup.com/qje/article/137/1/1/6383014 by guest on 10 March 2022 (see Online Appendix Figures CD1 and CD3). Along with education, gender is thus one of the only two variables in our data set for which a complete reversal of electoral divides seems to have taken place. 31 5. Other Socioeconomic Cleavages. Finally, our data set also makes it possible to study the evolution of the vote by union membership, public-private sector of employment, and homeownership. Union members have always been more likely to vote for social democratic and affiliated parties than nonunion members, although this gap has slightly declined in most Western democracies since the 1960s (see Online Appendix Figures CF5 (before controls) and CF6 (after controls)). This is also the case for public-sector workers and homeowners, who have remained more supportive of social democratic and affiliated parties than other voters in the past decades. 32

VI. CONCLUSION
The new historical database on political cleavages in 21 Western democracies introduced in this article reveals some important facts. In the early postwar decades, social democratic and affiliated parties represented the low-education and the low-income electorates, whereas conservative and affiliated parties represented high-education and high-income voters. These party systems have gradually evolved toward "multiconflictual" or "multi-elite" party systems in most Western democracies, in which higher-educated voters vote for the "left," whereas high-income voters still vote for the "right." 31. Several explanations have been given to this reversal. In the United States and Western Europe, the decline of marriage, the rise of divorce, and the economic fragility of women have been shown to be important drivers behind the emergence of the modern gender gap (Edlund and Pande 2002;Abendschön and Steinmetz 2014). In Northern Europe, the expansion of women's employment in the public sector has also been an important factor behind the increase in the vote for the left among women in recent decades (Knutsen 2001; we reproduce this result in Online Appendix Figure CD4). Women have also been more attracted by environmental issues, which have spurred women's support for green parties, while antiimmigration parties have generally found greater support among men (Givens, 2004; see Online Appendix Figures CD5 and CD6).
32. See Online Appendix Figures CF7 (before controls) and CF8 (after controls) for the sectoral cleavage and Figures CF9 (before controls) and CF10 (after controls) for support for left-wing parties among homeowners.
Results combining our database on political demand with political supply data from the CMP suggest that the emergence of a new sociocultural axis of political conflict has been tightly associated with the reversal of the education cleavage in Western democracies. As parties have progressively come to compete on sociocultural issues, electoral behaviors have become increasingly clustered by education group. This transformation has been most pronounced in democracies where parties compete most fiercely on this new dimension of electoral divides.
The divergence of political conflicts related to income and education documented in this article, two strongly correlated measures of socioeconomic status could also contribute to explaining why rising income and wealth disparities have not led to renewed class conflicts. It might shed light on the reasons growing inequalities have not been met by greater redistribution in many countries, as political systems could come to increasingly oppose two coalitions embodying the interests of two kinds of elites.
Although multiple lessons have emerged from this new database, we acknowledge that the analysis remains insufficient and is not exempted from limitations. First, the indicators of political supply used and more generally the CMP data capture the tendency of parties to emphasize specific issues and are therefore unable to perfectly measure their position on these issues. Moreover, the policy categories coded in the CMP database unfortunately remain very broad, which precludes us from analyzing in greater detail more specific types of issues such as gender equality, immigration, trade protectionism, or education policy. Addressing these shortcomings would require going back to the original manifestos and deriving new indicators from text analysis or alternative coding techniques.
Second, although our descriptive analysis has provided suggestive evidence that the reversal of the education cleavage and the rise of a new sociocultural axis of political conflict were interrelated phenomena, much remains to be understood when it comes to the mechanisms underlying this transformation. In particular, it remains unclear whether the reversal of educational divides was driven by a change in political supply independently from the structure of collective beliefs or whether shifting supply was on the contrary driven by changing social attitudes across education groups. While some studies have suggested that social divides between groups have remained stable on a number of issues in the long run (e.g., Evans and Tilley 2017; Downloaded from https://academic.oup.com/qje/article/137/1/1/6383014 by guest on 10 March 2022 Bertrand and Kamenica 2018), which would point to the role of shifts in supply, the data at our disposal does not allow us to disentangle these different channels of causality. A promising avenue for future research lies in establishing more directly the causal impact of political supply on the transformation of political cleavages. This would require identifying quasi-experimental settings in which parties exogenously change position on specific issues or suddenly move to emphasizing new concerns.
Finally, the electoral surveys exploited in this article rely on samples of a few thousands of voters available since the end of World War II that are sufficient to reveal major trends at the national level, but prevent us from carrying out more refined and long-run analyses. Other sources and methods, such as localized election results linked to census data, could be mobilized to broaden the historical perspective and perform more granular analyses.
All of these issues raise important challenges that we hope will contribute to stimulating new research in these multiple directions.