The Political Economy of the Prussian Three-Class Franchise

Did the Prussian three-class franchise, which politically over-represented the economic elite, affect policy-making? Combining MP-level political orientation, derived from all roll call votes in the Prussian parliament (1867–1903), with constituency characteristics, we analyze how local vote inequality, determined by tax payments, affected policy-making during Prussia's period of rapid industrialization. Contrary to the predominant view that the franchise system produced a conservative parliament, higher vote inequality is associated with more liberal voting, especially in regions with large-scale industry. We argue that industrialists preferred self-serving liberal policies and were able to coordinate on suitable MPs when vote inequality was high.


Introduction
Economic development is strongly influenced by the institutional framework (see, e.g., North, 1991). Suffrage institutions historically excluded large parts of society from political influence and arguably affected long-run development through biased tax systems and public goods provision (Engerman and Sokoloff, 2005). Concentration of political power in the hands of a small economic elite alongside a vast majority of the population without effective rights has been associated with the adoption of policies that mainly benefit the elite (Acemoglu et al., 2001(Acemoglu et al., , 2002. On the eve of WWI, Prussia was the last major country in the industrialized world to maintain a franchise that was unequal, non-secret, and indirect. Since 1850, members of the Prussian parliament were elected under the so-called three-class franchise system, which translated tax payments into voting weights. Voters were grouped into three classes (Abteilungen) by total direct taxes paid. Each of the three classes represented one third of the total direct tax revenue and received the same number of votes. Classes were filled with voters ranked according to their tax payments, starting from the highest, adding taxpayers until one third of the total tax base was reached, thereby allocating less than 5% of voters into the first class, on average. By providing economic elites, i.e. the landed nobility and industrialists, with more voting power, the system arguably produced policies tailored to fit this electorate (see Gerschenkron, 1966). 1 German economic historians widely agree that the franchise system was designed to "generate Conservative votes", 2 and thereby produced (socio-) economic policies favored by agrarian elites (see, e.g. Nipperdey, 1992, p. 86 andBurhop, 2011, p. 23). Kühne (1994a, p. 25) stresses that the system was "the bulwark of reactionary Prussianism par excellence." 3 Several attempts to introduce an equal voting system were blocked by MPs representing voting districts were the landed elite was concentrated (Ziblatt, 2008).
Despite this univocal assessment, the question whether the unequal distribution of voting rights in Prussia indeed produced conservative policies is ultimately an empirical one and has not been answered thus far.
Our paper fills this gap and investigates whether the three class franchise contributed to the adoption of conservative (socio-)economic policies by the Prussian parliament. Using newly digitized data from Prussian parliamentary minutes and voting censuses, we estimate the relationship between vote inequality across electoral constituencies and the political orientation of MPs. Our analysis of political orientation is based on all 329 roll call votes (RCVs) that took place in a period of more than 35 years during 1867-1903 (i.e. election periods 10 to 19). 4 This dataset allows us to capture the political economy during a period that is characterized by the industrialization of the economy and the secularization of society in Prussia and the German Empire more broadly. 5 Our research design allows us to analyze econometrically whether Prussia's three-class voting system systematically favored the adoption of certain types of (socio-)economic policies.
We find that MPs from constituencies with higher inequality, i.e., where a small number of affluent voters holds a large share of the political power, vote significantly more for liberal policies. We first show that a higher vote inequality is associated with a higher probability to vote for liberal positions in a selection of RCVs about taxation and public goods provision that had important implications for the Prussian economy. In our main analysis, where we use the universe of RCVs, we then relate vote inequality to an aggregate measure of MP-level political orientation and show that higher vote inequality is significantly associated with a more liberal orientation. While the three-class franchise may have resulted in a conservative parliament on average, our econometric evidence suggests that, within the Prussian parliament, vote inequality did not bias voting behavior of MPs into a conservative direction.
The results are robust to the inclusion of a sizable set of constituency-level control variables capturing regional development and thereby the political preferences of the median citizen. The set of controls includes a measure of landownership inequality, which itself generates more conservative voting. Complementing the analysis with biographical information for all MPs, we also inspect whether voting was biased toward MPs' personal agendas, potentially reflecting elite preferences. The inclusion of both sets of variables does not change our main qualitative findings. Furthermore, although we find that party affiliation naturally captures a lot of the variation in political orientation of MPs, there is sufficient within-party variation in political orientation so that our main results hold conditional on party fixed effects. 6 4 The starting point is chosen because in January 1867 Prussia reached its maximum extension after incorporating Schleswig-Holstein. The choice of the end year is due to the fact that, from election period 20, there was a change in the procedure to elect MPs, and in 1906 a range of constituencies were split. 5 The period coincides with the transition from agriculture to industry (see, e.g., Becker et al., 2011), the move from liberalism to protectionism (see, e.g., Lehmann, 2010b), the secularization of society (see, e.g., Becker and Woessmann, 2013;Becker et al., 2017), the introduction of a welfare state (see, e.g., Bauernschuster et al., forthcoming;Lehmann-Hasemeyer and Streb, 2017), and the demographic transition (see, e.g., Becker et al., 2010). 6 Our main specification pools the voting behavior of MPs over election periods. Shifting majorities and the famous conservative turn in German politics during the 1880s may raise the concern that our results are driven by votes cast during the liberal period until 1880. Thus, we estimate the relationship of interest by election period to examine changes over time. Consistent with the notion of a largely stable policy arena during the entire period, results show substantial stability over time. Major economic changes that took place during the period seem to be not reflected in the political economy of voting in the Prussian Parliament.
To rationalize our results, we argue that the elite prefers policies which are in their own interest. Among the elite, large landowners prefer conservative policies that maintain their social and economic status whereas large industrialists prefer liberal policies that allow them to reap the benefits of increasing economic activity in the industrial sector. The rise of large-scale industry in late 19th-century Prussia increased earnings and tax contributions of industrialists. This development resulted in a stronger income dispersion and therefore in higher vote inequality in industrial regions than in agricultural regions where the income of landowners remained largely unchanged (Kuznets, 1955). As a consequence, the highly unequal distribution of individuals across voting classes in industrial regions, created a small and homogeneous first class in which industrial elites could coordinate on MPs who supported liberal policies. Indeed, we provide evidence that vote inequality is associated with more liberal MPs only in regions with a high share of large industrial firms.
To measure political orientation of MPs, we digitized the universe of RCVs from the Prussian House of Representatives in the period 1867-1903. We use Poole's (2000) optimal classification (OC) method to compute 'ideal points' that locate more than 1,900 MPs in a multi-dimensional policy space according to their voting behavior in multiple RCVs across election periods. We discuss details of the method below, but would like to note here that the OC method resembles a factor analysis and follows a standard approach to gauge political preferences, summarized by Clinton (2012), that is used by political scientists but also by economists and economic historians. 7 The agnostic OC algorithm correctly predicts ca. 95% of MP voting behavior using a two-dimensional policy space. Based on knowledge about historical party platforms, we interpret these dimensions to reflect an economic "liberal-conservative" and a cultural "secular-religious" orientation. In the main econometric analysis, we separately regress each of the two ideal-point coordinates on our measure of vote inequality, which is a Gini coefficient computed from data on the constituency-level distribution of voters across the three voting classes.
We expect only the "liberal-conservative" dimension to be influenced by vote inequality and use the "secular-religious" dimension in a placebo test, because we do not expect it to respond to vote inequality. Conflicts between Protestants and Catholics that culminated in the secularist policies of the Kulturkampf ('culture struggle') era in the 1870s transcended class structures and should not be reflected in the distribution of voting power.
Indeed, we find that vote inequality does not predict for which type of cultural policies MPs vote. Instead, the cultural orientation of MPs can almost exclusively be explained by the religious and ethnic heterogeneity of the electoral constituencies.
We aim to mitigate endogeneity concerns in the relationship between political orientation and vote inequality using the share of voters in the first class as an instrumental variable. The allocation of voters to the first class reflects a local threshold between terciles of the tax base that are arbitrary from a national perspective, as we will explain in more detail below. In locations with a large number of high income earners, voters with high incomes are more likely to be allocated to lower classes than elsewhere. After conditioning on average tax payments, the remaining variation thus reflects arbitrary differences in the number of voters in the first class across districts. Resulting IV estimates confirm the link between vote inequality and liberal orientation.
Our findings are in line with a recent literature debating the role of elites in the process of modernization and democratization. This literature presents results that are broadly consistent with the idea that elites prefer 'liberal' policies if in their own interest. 8 For example Ashraf et al. (2017) argue that capital-owning elites in industrializing Prussia granted more freedom to the masses to incentivize their labor effort. In Tsarist Russia, liberal nobles cooperated with the peasantry to expand public goods provision (see Nafziger, 2011). In 19th-century Sweden, a suffrage reform shifted voting power from landed elites to industrialists, resulting in higher investments in railways and structural change (see Hinnerich et al., 2017). In Revolutionary France, the enlightened elite supported more 'liberal' education policies (see Squicciarini and Voigtländer, 2016).
We also contribute to a large literature that uses RCVs to study the orientation of historical parliaments (see e.g. Schonhardt-Bailey, 1998;Rosenthal and Voeten, 2004; Heckelman and Dougherty, 2013). We are not aware of work that systematically analyzes RCVs from the Prussian House of Representatives. However, inspecting a period that strongly overlaps with ours, Häge (2018) provides a descriptive analysis of the policy space across different Reichstag legislatures during the Bismarck era in the German Empire. 9 While their orientation shows similarities, the two parliaments differ in their franchise system. Thus, discrepancies in dimensionality and party affiliation maybe the outcome of a fundamentally different political process. In addition, our paper focuses on explaining rather than describing the orientation of parliament.
Earlier work that tried to explain voting behavior in the Prussian House of Representatives has typically focused on selected RCVs debating important issues. Ziblatt (2008) analyzed one RCV seeking to abolish the three-class franchise in 1912 and finds higher landownership inequality but not income inequality to be associated with voting in favor 8 The competing hypothesis holds that elites adopt policies in favor of the lower classes because of a prevailing threat of revolution by the masses (see Acemoglu and Robinson, 2000;Aidt and Franck, 2015). On the other hand, Aidt and Jensen (2017) argues that the introduction of the secret ballot in the U.S. and Europe came as a byproduct of the process of modernization due to the decline of social control and vote buying. 9 The dimensionalities of the Reichstag during the later Weimar period and the earlier Frankfurt National Assembly are depicted by Hansen and Debus (2012) and Herrmann and Sieberer (forthcoming), respectively. of preserving the three-class system. 10 Mares and Queralt (2014) analyze one RCV on the introduction of a Prussian income tax system in 1891, showing that higher landownership inequality is associated with voting for the adoption of an income tax that disproportionately burdened the industrialists. 11 We go beyond previous work by using the universe of RCVs in the Prussian House of Representatives during 1867-1903, allowing us to characterize the political economy of Prussia in this crucial period of economic development.
The remainder of the paper is structured as follows. Section 2 elaborates on the institutional context of the Prussian three-class franchise, roll call voting in parliament, and the political landscape. In Section 3, we derive a conceptual framework to explain our main findings. Section 4 introduces the regression framework and the data which we use to estimate the relationship between vote inequality and political orientation. Section 5 introduces our measure of political orientation based on Poole's (2000) optimal classification method and its application to the Prussian House of Representatives. Section 6 presents the results and our interpretation. Section 7 concludes.

Historical context
In 1848/49, Prussia introduced a bicameral parliament, with an appointed House of Lords (Herrenhaus) and an elected House of Representatives (Abgeordnetenhaus). The constitution, imposed by the King, stipulated that the lower chamber was elected by male citizens aged 24 or older who had paid taxes in the preceding year. Voters were allocated to three classes reflecting terciles of the total direct tax base. Votes were effectively weighted in proportion to payments of direct taxes (class-tax, classified income tax, real estate and property tax, and business tax): the upper tercile of tax payments had as many votes as the middle tercile and the lower tercile. On average, first-class voters had about 17.5 times the number of votes of third-class voters, obscuring even stronger inequality at the local ward (Urwahlbezirk) level. In the extreme, a single taxpayer who paid a third or more of direct taxes in a ward would have as many votes as all of the taxpayers in the lower tercile combined. 12 The thresholds for allocating voters into the three classes according to their tax 10 Lehmann-Hasemeyer et al. (2014) show that the Berlin Stock Exchange reacted negatively to the extension of the franchise in Saxony in 1896 and 1909. 11 Debus and Hansen (2014) use several Weimar Republic roll calls to study the voting behavior of women in parliament. Lehmann (2010a) analyzes several RCVs from the Reichstag during the Weimar Republic to analyze how coalitions formed between MPs of different parties. 12 The most prominent example was the industrialist Alfred Krupp who was the only voter in the first class in his ward. However, the pattern can also be found in rural wards where the income distribution was dominated by a local noble estate owner. payments were determined at the municipality level. 13 Municipalities with more than 1,750 inhabitants were subdivided into wards but maintained the municipality-level tax thresholds. The mayor was responsible for the division into wards, frequently leading to gerrymandering (Kühne, 1994a;Richter, 2017). In cases where social stratification led to sorting into city quarters, the first and second class remained unpopulated in poorer wards, whereas the first class was populated by many voters in richer wards, reducing individual voting power in wealthy city quarters. 14 Elections proceeded in two stages: in stage one, the primary elections (Urwahlen), voters elected electoral delegates (Wahlmänner). In stage two, the electoral delegates elected the members of the House of Representatives. Stage one was voting by a show of hands, i.e. non-secret, and took place at the level of wards. 15 Each class in a ward elected the same number of delegates, ranging between 3 and 6 per ward, one per 250 inhabitants (according to the last census). At the second stage, electoral delegates from all wards in an electoral constituency met to elect between one and three MPs, depending on its population size. Electoral constituencies comprised at least one, but typically several (administrative) counties (Kreise). Put differently, the approximately 500 Prussian counties are nested in 252 electoral constituencies, allowing us to map aggregated county-level census data to MPs representing these constituencies. 16 The three-class franchise was frequently disputed but persisted in parallel with the equal voting system of the Imperial Diet, the Reichstag, until 1913.

Roll call voting
The Prussian constitution of 1850 stipulates in article 64 that the King and each of the chambers have the right to introduce bills. In the House of Representatives, legislative proposals can be introduced by the President of the chamber or by groups of at least 10 MPs. Proposals introduced by the King or the other chamber are voted on by absolute 13 In cases when several smaller municipalities were combined into one ward, thresholds were determined at the ward level.
14 The tax reform of 1891 and the subsequent voting reform for the 1893 election introduced classification of voters according to ward-level thresholds. This led to the arbitrary upgrading of low income voters into the first and second class in poor quarters and downgrading of high income voters into the second and third class in wealthy quarters. According to the most prominent example, the Chancellor of the German Empire, Bernhard von Bülow, was required to vote in the third class in 1903, because he lived in the Reich Chancellery which was located within the ward with the highest tax base in Berlin. 15 Because voting in primary elections was not secret, voters in the upper tercile, who often were employers of those in the lower terciles, held coercive power. Thus, even if we had information on voting behavior by the third class in primary elections, they would not necessarily reveal true preferences. Intimidating behavior by industrial employers, rural landlords, and public officials has been documented for elections to other contemporary German parliaments including the Reichstag and the Prussian House of Representatives (see Kühne, 1994a;Mares, 2015;Ziblatt, 2009). 16 Unfortunately, since MPs do not represent a county, we cannot match MPs to specific counties, even though our data is available at this level. majority. To vote, the majority of members of a chamber must be present. To adopt a law, all three sides (King and both chambers) must agree, so they are fully equal in the legislative process.
Voting in the House of Representatives took several forms. When the majority on an issue was clear-cut, a count based on standing or sitting MPs established the result. RCVs were triggered in one of two ways: first, upon request of at least 50 MPs (see Plate, 1903, p. 187;§61 GO). Second, with a majority of less than 15 votes, a roll call could be requested without further support/signatures (see Plate, 1903, p. 187;§58 GO). From 13th February 1875, roll call votes were replaced by a division of the assembly (Hammelsprung). 17 This reform likely explains why the number of roll call votes declined after 1875 (see Table B.1).
The absolute majority decided on all votes. Abstentions were typically not included in the calculation of the majority. 18 Since RCVs were only called on close and/or contentious issues, they are likely to be quite selective (see Carrubba et al., 2006). This has advantages and disadvantages: RCVs are associated with highly important issues, which is a clear upside for our analysis.
On the other hand, party cohesion is likely to be higher in RCVs (see Yordanova and Mühlböck, 2015). This implies that if we could observe all votes in parliament, the actual variation in political orientation across MPs would likely be larger than what we observe using only RCVs. Against this backdrop, the fact that we do find substantial variation in political orientation, above and beyond what is predicted by party membership, is striking and makes us confident that we are measuring relevant information.

The political landscape
The Prussian party landscape stabilized during the early periods of our analysis. Political fractions and parties formed during the 1860s, and gained relevance after the German Reichstag was established in 1871.
Drawing on biographical information from Kühne (1994b), we report the party affiliation of MPs in Table 1. 19 The table depicts an aggregate number of 3,658 MP-by-period observations of party affiliation between 1867 and 1903. A range of smaller factions were represented in parliament. We classify members of smaller factions into larger party groups as indicated in columns 2 and 3. This leaves us with six main parties which were active 17 In the German context, MPs would (re-)enter the House via different doors to facilitate counting. 18 In our analysis period, there were 4 exceptional cases where abstentions were, however, included in the count (Plate, 1903, p. 180). 19 In some cases when party affiliation was uncertain, we used the most likely party affiliation given in Kühne (1994b). In cases of changes in party affiliation within a legislature, we use the initial party affiliation. Party changes within legislature are extremely rare and affect well below 1% of MPs. In the pooled analysis, in cases when MPs changed party affiliations between legislatures, we assigned the party for which the most roll call votes were cast. during our period of analysis (and the residual category 'Other' is mainly populated by liberal MPs).
Parties and their members strongly overlap with the German Reichstag. 20 Platforms covered a spectrum of left and right wing policies. However, due to the inequalities embedded in the franchise system, platforms addressing social issues and redistribution such as the Socialist Workers' Party did not manage to obtain any seats in parliament until 1908. 21 Parties are clearly distinguishable in their religious views as well as in their support of the Imperial Government, in particular the political agenda of chancellor Otto von Bismarck (see Treue, 1975 20 More than 30% of Prussian MPs were members of both parliaments at some point in time. Concerns that MPs who belonged to both parliaments were systematically different can be addressed with a dummy variable for dual membership. In unreported regressions, we find that adding such a control variable does not change our results. 21 In that sense, the Prussian Parliament as a whole was a conservative one. But as argued before, our interest is in the link between vote inequality and political orientation within the three-class franchise.

Conceptual framework
To better understand how vote inequality affects the political orientation of MPs, this section discusses the determinants of variation in vote inequality. The three class franchise was created at a time when industrialization was still in its infancy and income inequality was relatively low. In the absence of large-scale industry, large landowners paid a sufficient amount of taxes to populate the first two classes, despite the fact that they enjoyed a range of tax exemptions and paid little taxes on land-their main source of income. Due to the rapid rise of large-scale industry in the second half of the century, industrialists started to obtain high returns to capital investment and incurred an increasing share of the total tax base. In industrializing regions, this development lifted the new industrial elites into the first two classes, where they crowded out parts of the traditional elites. Thus, the rise of industry created an unforeseen regional concentration of high incomes and thereby a higher level of vote inequality in industrializing regions while vote inequality in agricultural regions was less affected. This description is in line with the developments observed by Kuznets (1955) who also relied on Prussian income data (also see Grant, 2002).
Under the assumption that owners of large firms favor liberal policies that allow them to explore economic opportunities, the new concentration of industrialists at the top of the income distribution led to the creation of a small first (and likely second) class with liberal preferences. Furthermore, under the assumption that landowners favor policies that protect their (socio-)economic status, in rural regions with little competition over entry into the upper classes, political preferences in the first and second class remained conservative.
Next to these polar cases, the growth of industry in agricultural regions may have increased heterogeneity in political preferences and thereby resulted in coordination issues within the first class leading to the election of less conservative MPs. These plausible assumptions can explain how variation in vote inequality arose and why higher vote inequality may be associated with the election of more liberal MPs, especially in districts with large-scale industry. 22 The assumption that industrial elites prefer liberal policies is supported by the fact that, among the three types of liberties-individual, economic and political-liberal parties in late 19th-century Prussia are most strongly associated with economic liberalism (see Mork, 1971). Parties self-identifying as 'liberal' in their name propagated liberal economic policies, stressed the need for (domestic) competition, and supported the separation of church and state. Considering the dynamism of Prussia's second phase of industrialization (see Becker et al., 2011), we expect industrialists to be particularly keen to grasp arising economic opportunities. On the other hand, higher levels of landownership concentration may reflect higher demand for social control due to a higher risk of expropriation for the elite (Weber, 1917;de Tocqueville, 1856). Since landowners rely on an immobile factor of production, they prefer conservative policies that sustain inequality (Acemoglu and Robinson, 2000;Boix, 2003). As shown by Ziblatt (2008), MPs from Prussian regions with higher landownership inequality voted to preserve the three-class franchise. Our empirical results, derived from RCVs, will indeed show that MPs belonging to liberal and conservative parties are systematically sorted over a spectrum of voting behaviors from liberal to conservative. 23 The assumption that a stronger concentration of elites in the first class facilitates coordination on a suitable MP is supported by qualitative evidence about the functioning of the two-stage election system. First, contemporary sources suggest that electoral delegates often served multiple times in a row and gained some importance as de facto "electoral delegate patricians" (see Kühne, 1994a, p. 52). 24 A more concentrated economic elite will find it easier to influence electoral delegates, 25 and in extension local MPs. 26 Second, industrial elites were able to coordinate on candidates more easily because of higher population density and lower information costs in urban areas (see Kühne, 1994a, p. 51-52). In fact, urban areas had better infrastructure which made it more likely for urban elites to interact and to be able to meet electoral delegates and local MPs and thereby to influence their selection and roll call voting behavior. To the extent that urban elites are industrial elites, both of these mechanisms lead us to expect a stronger link between vote inequality and liberal voting in areas with industrial elites. 27 Furthermore, during the second phase of industrialization, industrial elites may have formed a strategic coalition with the third class to support policies that are mutually beneficial for capitalists and workers (also see Galor and Moav, 2006;Doepke and Zilibotti, 2008;Llavador and Oxoby, 2005). Mutual benefits -e.g. from voting in favor of institutions that promote freedom of occupational choice and generate a more skilled labor force -cre-23 This argument is by no means circular: liberal parties stress economically liberal policies in their party programs and their collective roll call voting behavior is consistent with these purported policy aims. 24 The two-stage election process with electoral delegates resembles the US context. Interestingly, research for modern-day US finds, using measures of state party positions, that low-income preferences are underrepresented already in parties' campaign appeals, i.e. at a very early stage in the policymaking process. This effect is more pronounced in states with greater economic inequality (see Rigby and Wright, 2013). 25 This is supported by Kuehne who suggests that MPs regularly met with electoral delegates as intermediaries between the electorate and MPs. 26 An additional contributing factor shown in modern-day cross-country data is that higher levels of income inequality depress political interest and participation in elections among all but the most affluent citizens (Solt, 2008). This is not the case in our Prussian data, where there is no statistically significant link between vote inequality and turnout of the third class. 27 It is also interesting to note that liberal parties have higher shares of MPs that are industrialists, as we discuss in Section 4.2. ate a strategic complementary between the first and third classes at the cost of conservative voters. Such a complementary between capital and labor likely matters more for owners of large firms that particularly benefit from division of labor between specialized workers.

Model and data
To estimate the relationship between the political orientation of MPs and the vote inequality in their electoral constituency, we generate a dataset in which the unit of observation is an MP, or more precisely, MP-by-constituency. The reason for the subtle distinction is that a small number of MPs (104 cases) represent different constituencies over their career. 28 Our results are robust to dropping such MPs altogether and instead to only using MPs who represent the same district throughout all of their career.
We employ the following standard OLS framework: where PolId ic constitutes a measure of political orientation of an MP i in constituency c. The political orientation of an MP will be measured in two ways: a) using the 'yea/nay' vote of a single RCV or b) using aggregate scores derived using the optimal classification algorithm described in Section 5 that locates the voting behavior of MPs across all RCVs into a multi-dimensional policy space. Our main analysis will be based upon the two scores that constitute the ideal points of each MP in policy space. 29 VoteIneq c is our measure of vote inequality, calculated from the distribution of voters across the three classes. X c is a vector of constituency-level characteristics, Z i is a vector of individual-level MP characteristics.
Constituency-level control variables include standard measures of (land-ownership) inequality and development as used in the literature (see below). These can be aggregated from county level data. Characteristics varying at the level of individual MPs are personal characteristics, such as age, religion and occupation, but also electoral turnout and margin of victory in the election in which the MP was elected, and party affiliation. The latter is not a regressor like any other: by virtue of parties' ability to draw like-minded representatives 28 MPs switching constituencies throughout their career may be a highly selected group that strategically ran in various electoral constituencies. In (unreported) regressions we find their voting behavior to not respond to local constituency characteristics. 29 When simply swapping the two scores as dependent variables, we assume that the standard errors are uncorrelated across equations. In fact, since the correlation between the two dimensions is approximately 2%, this seems reasonable. However, using the potentially more efficient seemingly unrelated regression (SUR) method that allows standard errors to be correlated, we find very similar results.
to stand for them, we expect party dummies to absorb a substantial part of the variation in OC scores. Most variables are drawn from county-level data provided by Galloway (2007) and iPEHD (Becker et al., 2014). As constituencies combine up to three individual counties, we aggregate county level data to the constituency level. Figure 2 depicts the spatial variation in vote inequality with darker colors reflecting higher inequality. Vote inequality is a Gini coefficient calculated from newly digitized county-level distribution of voters across the three classes. 30 Using the fact that each class contributed exactly one third to the tax base, we calculate the Gini coefficient based on the number of voters in each class. Notice that vote inequality assigns the same power to each voter in the same class so that we do require information about the complete tax contributions in a constituency. 31 Our main analysis will use the distribution of voters across classes from the election of 1893 reported in Königlich Preußisches Statistisches Bureau (1864Bureau ( -1905, the first time such data was published for Prussia in its post-1866 borders. 32

Constituency characteristics
Following the literature (Lehmann, 2010b;Mares and Queralt, 2014;Ziblatt, 2008), we include a set of standard explanatory variables for MP voting behavior in the Prussian House of Representatives in our vector of constituency characteristics. To compare the impact of vote inequality with other forms of inequality, we include landownership inequality as a measure of immobile wealth inequality in our estimations. 33 We expect this variable to correspond to a more conservative orientation. Other control variables include the industrial employment share, the urbanization rate, the share of Protestants, the linguistic fractionalization, the share of the population that lives in their municipality of birth, as well as the literacy rate to capture heterogeneity in the available stock of human capital. 34 These indicators measure differences in development, cultural heterogeneity, and structural differences across regions. 30 We aggregate county-level data to the constituency-level before calculating the Gini. 31 Different from a measure of income inequality, vote inequality ignores within-class variation in income. Most likely vote inequality is correlated with income inequality because variation between income groups is a substantial part of the overall variation in incomes. Since this paper is interested in the franchise system, we use the vote inequality Gini that exactly captures the different degree of representation enjoyed by the three classes. 32 Our analysis will also test robustness of the results using the distribution of voters across classes from the election of 1861. These data are less prone to potential reverse causality but are unavailable for the new provinces annexed by Prussia after 1866, i.e. we lose more than 350 MPs. 33 The correlation between vote inequality and landownership inequality is approximately 69%. 34 For further information see Appendix A.
The period under analysis is a critical juncture in the transition from agriculture to industry. At the beginning of the period, slightly less than half of the population was based in agriculture. At the end, agricultural employment made up merely one third of the labor force. However, these structural changes may be an outcome of the political process and therefore prone to concerns of reverse causality. Thus, in our regressions we include variables that are measured closest to the starting point of our analysis in 1867. 35 If an MP represents the average individual in his constituency, we expect local conditions to explain a large part of the variation of MPs' voting behavior. However, if MPs represent the local elite but also act in their own interest, individual characteristics may explain a lot of the variation. We can probe this by also controlling for MP characteristics. Catholic MPs are affiliated with the Centre party and the Polish minority party. We expect them to lean towards policies that support a strong position of the church.

MP characteristics
Additional attributes that are available from MP biographies include nobility status, academic titles, residence in the electoral constituency, and retirement status. 40

Political characteristics of MPs
Four variables broadly account for the political characteristics of MPs and their relationship with their constituency. RCV participation measures the share of the total RCVs in a period the MP participated in. This variable may be related to both representation and political orientation. Electoral margin arguably captures the strength of the mandate and whether the MP seat was strongly contested. The literature (see e.g. Fowler, 2005;Fowler and Smirnov, 2005) argues that MPs are more likely to enforce extreme versions of their political agenda if the margin of victory is large. Table A.3 in the Appendix shows an increasing average margin across periods, reflecting incumbency effects and an increasingly static policy arena around the turn of the century.
Electoral turnout is argued to increase in the level of representation across electoral systems and thus may affect our estimates (see, e.g. Jackman, 1987). Importantly, our turnout measure reflects participation of electoral delegates (Wahlmänner) in the second stage of the electoral process and not the turnout of voters in the primary elections (Urwahl). 41 The number of MPs reflects so-called multi-member constituencies that have been argued to foster electoral competition and broader coalition building, leading to more redistributive policies (Persson and Tabellini, 2004).

Measuring political orientation using Optimal Classification
To understand whether the franchise biased parliamentary voting patterns, we inspect MPs' overall voting behavior instead of voting in single roll calls or changes in voting behavior over time. 42 This requires us to classify all votes into categories that reflect a political orientation. Manually classifying votes from 329 roll calls into liberal and conservative (or 40 The variable 'noble title' is unity if the affix of an MP includes the term Freiherr, Graf, Baron, Edler Herr, Prinz, von, vom, am, aus dem, de, du, van, or zur. The variable 'academic title' is unity if the title of an MP includes the term Dr. or Dipl. Ing. 41 The turnout in this second stage is about 90% on average and remains stable across periods. For the smaller sample of constituencies in the pre-1867 borders, we also have information on the turnout in the primary elections by class. As expected, turnout increases in the level of representation (first class: 57%, second class 44%, third class 24%. However, neither of these measures is significantly related to political orientation. 42 Since the franchise system was not subject to change during our period under investigation, we do not expect major changes in voting behavior over time. Neither procedural rules on roll call voting nor on plenary agenda setting experienced fundamental changes during the period. into secular and religious) leanings is a task that would be prone to many arbitrary decisions.
Simply reading minutes of parliamentary debates preceding RCVs, we could gain a nuanced impression of speakers positions but still incorrectly classify MP positions because debates do not reveal underlying rationales of voting choices. Furthermore, classification requires the strong assumption that all MPs perceive the political dimension underlying a vote in the same way. Even a selected subset of RCVs may lead to erroneous classification.
Suppose we selected a set of RCVs that evolve around secular-religious motives during the Kulturkampf, including a vote on a bill that exempts priests from punishment for reading mass and administering the sacrament. It will be hard for us to classify whether an MP's vote is motivated by his allegiance to the Pope, his support of Bismarck, or his liberal orientation that rejects state intervention. Votes may indeed reflect multiple motives and dimensions that are obscured. Manually determining how to code 'yea' or 'nay' votes in any RCV requires us to make assumptions about the position of all MPs relative to the matter.
To avoid the pitfalls of 'manual' classification, political science has developed scaling methods that are widely used in the analysis of parliamentary voting. In our main approach, we follow this literature and analyze RCVs using the non-parametric optimal classification (OC) method introduced by Poole (2000). 43 Similar to a principal component analysis, the OC method extracts one or more latent variables from the votes that are fed into the algorithm. This approach is superior to a researcher's classification because decisions will not be biased by prior knowledge (and interpretation) of the content of the RCVs or the party affiliation of MPs. The algorithm does not rely on an ex ante classification but will allow an ex post interpretation of the latent dimensions it uncovers.
OC is a non-parametric scaling procedure that classifies a matrix of yea and nay votes. 44 After generating an 'agreement matrix' between MPs, the algorithm extracts coordinates that locate MPs within a policy space using a pre-specified number of dimension.
For simplicity, suppose that voting in parliament is exclusively driven by one dimension -liberal or conservative motives. The OC method interprets votes of MPs as revealed preferences from a choice between a proposed policy and the status quo. The underlying assumptions are that legislators have Euclidean preferences defined over the policy space and that they vote sincerely for the alternative closest to their 'ideal point '. 45 Since neither 43 This methodology is related to the W-NOMINATE method that is frequently used in economics and has been adopted by the media when discussing the U.S. parliament. According to Rosenthal and Voeten (2004), the OC method is superior to parametric methods such as W-NOMINATE, when analyzing a multi-party setting such as ours. 44 For a more technical discussion, see Poole (2000). 45 In fact, following Hix and Noury (2016) we should call these 'revealed positions' rather than 'ideal points' to acknowledge that their position is not solely driven by personal preferences but also by institutional factors the exact position of the proposed policy nor the status quo is known, we only observe two positions reflecting 'yea' or 'nay' votes. These two positions can be intersected by a cutting line separating groups of legislators with different positions. To illustrate this procedure, we discuss the policy space and the cutting lines for two selected RCVs in Appendix Section C. Using a larger number of RCVs allows to estimate MP positions more precisely due to varying distances between ideal points, proposal, and the status quo. The OC algorithm will produce a rank ordering of MPs and vectors of relative distances that constitute the latent dimensions.

OC results for the Prussian parliament
Our main analysis pools all 329 RCVs across ten election periods providing us with substantial variation in political orientation within parties. Each MP representing the same constituency during one or several election periods constitutes one observation, similar to the 'fixed career model' as described in Asmussen and Jo (2016). The various issues voted upon over this long period are likely such that party discipline cannot be always imposed, giving us substantial within-party variation in political orientation. Pooling across periods also reduces the impact of outliers, and the resulting ideal points are more likely to represent an unbiased measure of constituency-level policy preferences. Furthermore, some variation derives from the fact that MPs switched their party affiliation over time.
We assign party affiliation based on the party for which an MP casts the highest number of votes. In an extension, we show that findings are qualitatively similar when inspecting election periods separately, thus allowing for party switching and relaxing restrictions on legislators' ability to move within the policy space.
We identified roll calls by working through the universe of parliamentary minutes.
For all RCVs, minutes list the name of MPs and record 'yea' and 'nay' votes as well as abstentions and absence. 46 We coded these votes and matched MPs by name and period of activity to the Kühne (1994b) dataset that provides us with biographical information and constituency identifiers. The quality of the match was manually checked using all available historical sources. Table 2 presents descriptive information generated when scaling the RCV data with the OC method for the ten election periods. The analysis is limited to legislators who voted such as the electoral system. The Prussian parliament is akin to a presidential system in which the political leader (i.e. Bismarck) builds issue-by-issue coalitions. Consequently, in line with models of spatial voting behavior, we can assume that the Prussian parliament resembles a floor agenda model without restrictions to agenda setting (Hix and Noury, 2016). 46 We are confident that we identified all RCVs in the period because the voting sequence of MPs systematically rotates over the alphabet restarting with the letter A after each election or recess. This procedure helped us ensure that we are not overlooking any RCVs. From 1883, RCVs are also listed in the index of the parliamentary minutes. in at least 15 RCVs and excludes extremely lop-sided votes where the minority consisted of less than 2.5% of MPs. These criteria apply to 328 of the 329 RCVs and to 1903 of 2291 MPs, amounting to a total of 107,282 individual votes (58,540 yea and 48,742 nay). 47 Abstentions and absence during an RCV are coded as missing values. Similar to Häge (2018) who analyzes the Reichstag during the Bismarck era, we find abstentions to be rare (487 votes) but absenteeism to be substantial. Table 2 also provides information on the correct classification of votes (columns 5 and 7), the Average Proportional Reduction in Error (APRE) (columns 6 and 8), and eigenvalues (columns 9 to 11) comparing a one-dimensional and a two-dimensional case. We correctly classify 89.4% of votes when assuming one dimension and can increase the accuracy of classification to 96% adding a second dimension in the pooled sample. Similarly, the two-dimensional approach substantially increases the correct prediction of votes in most of the individual election periods. Comparing APREs, we find that two-dimensional scaling substantially improves the error reduction by approximately 17 percentage points in the pooled sample and by more than 10 percentage points in all periods, except periods 10, 12, and 13. Similarly, eigenvalues fall off quickly in these periods suggesting that parliamentary debates during periods 10, 12, and 13 may be sufficiently described as developing alongside a single dimension, whereas all other periods are two-dimensional. 48 49

The policy space for the Prussian parliament
The graphical results of the optimal classification approach for our pooled dataset are presented in Figure 1a. Assuming a two-dimensional policy space, the OC procedure generates two coordinates that reveal the position of each MP within a unit circle, i.e. coordinates assume values between −1 and +1. Colors and letters indicate party membership as presented in Table 1. We find that MPs belonging to the same party are spatially clustered, 47 Changing the minimum vote requirements comes with a trade-off between the precise estimation of ideal points and the number of MPs who cast a sufficient number of votes. In one case, in election period 13, when only 15 RCVs took place, the number of MPs who participated in all 15 RCVs is too small to generate a meaningful policy space. Based on post-estimation inspections of the policy space for period 13, we reduce the minimum value to 12 votes, which provides sufficient information to grant the characteristic clustering of Prussian MPs into parties. 48 Eigenvalues are computed on the basis of the so-called agreement matrix between representatives (in election period t), i.e. the square matrix where entry (i, j) is the fraction of roll call votes where representatives i and j vote the same way (in election period t). The faster the eigenvalues fall off, the stronger the evidence that only a small number of dimensions is required to describe systematic variation in voting patterns. 49 Figure D.1 in the Appendix shows 11 corresponding "scree plots" following Cattell (1966). Figure D.2 in the Appendix shows the cutting lines for the pooled sample and by election period. The criss-crossing lines show how different roll calls bring together different groups of representatives on the Yea or Nay sides of the cutting lines. The angle of the crossing lines with respect to the horizontal axis gives us an indication of whether the issues voted on are more divided along the first dimension or along the second dimension. Another way to present the same information is by virtue of Figure D.3 in the Appendix which shows a histogram of cutting line angles in the pooled sample and by election period. reflecting the fact parties are groups of like-minded representatives. Since the OC algorithm is agnostic about the spatial orientation of its output, we anchor the policy space using leaders of the Conservative and Centre parties. 50 Using information about the political landscape in Prussia outlined in Section 2.2 we interpret these dimensions as a liberal-conservative and a secular-religious dimension, reflecting the main cleavages between parties in parliament. Even though parliament did not capture the full spectrum of political views (remember the absence of Social Democrats), it included parties advocating economic liberalism of the Manchester type and parties ad-  Furthermore, these MPs are almost exclusively located in the northern part of the policy space, opening up a second dimension. This resonates with the fact that the Centre Party is largely composed of Catholic MPs whose agenda was to defend the power of the 50 Our choice of the anchors in each dimension is guided by the research question (MPs standing for a certain 'orientation'), the length of tenure in parliament, and the number of votes cast in roll calls (for representativeness and stability of estimation). In the pooled analysis of all election periods from 1867 to 1903, we use one MP from the Conservative Party (Friedrich von Wedell-Malchow) to define a "conservative" orientation of the policy space, and one MP from the Centre Party (Ludwig Windthorst) to define the "religious" orientation of the policy space. In the election period-specific analysis, we use Wedell-Malchow and Windthorst for election periods 10 to 16, but for election periods 17 to 19 we use Friedrich Wilhelm Graf zu Limburg-Stirum as the "conservative" anchor and Clemens August Freiherr von Heereman von Zuydwyck as the "religious" anchor, because Wedell-Malchow and Windthorst died. All anchors were also members of the Reichstag. 51 To compare dimensionality in the German Reichstag across election periods during the Bismarck era, Häge (2018) rotates the policy space to align the Conservative Party with the 45 degree line arguing that legislators held stable conservative positions in both dimensions.
Catholic Church during the Kulturkampf. To ensure that we indeed capture the secularreligious orientation, which will be used in a placebo test, we fix the centroid of the Centre Party to the vertical axis. Comparing Figures 1b and 1c, the policy space rotates only slightly. While Centre Party MPs exclusively locate on the religious side of the secularreligious axis, MPs from the National Liberals and Free Conservatives can be found in the secular southern region. However, since the Conservative Party was divided over secularization and orthodox-Protestant MPs sympathized with the position of the Centre Party, we find considerable variation across liberal and conservative parties. Similarly, while many free liberal MPs reveal positions that support secularization, we find the majority to be neutral in this dimension neither favoring church nor state involvement in matters of education and civil society. 52 We confirm the graphical observations of a clear ranking of parties along the liberalconservative spectrum in OLS regressions presented in column 1 of Table 3 To test their relationship with the franchise, the rotated ideal points serve as dependent variables in the regression framework described in Section 4.

Results and interpretation
This section presents results from regressions of the econometric model described in Section 4. We will first show results based on voting in selected RCVs, then results based on the ideal points derived from the OC method (displayed in Figure 1) using the full universe of RCVs across election periods 10-19 (1867-1903), and finally results based on ideal points generated for an election-period-specific analysis. In all specifications we cluster standard errors at the level of constituencies. 54 52 To address the concern that the distribution of MPs across the policy space in the pooled sample is driven by specific periods that constitute outliers, we executed a 'jackknife' analysis, i.e. consecutively dropping each period to probe sensitivity. Results of this procedure are presented in Figure D.4 in the Appendix. 53 Results are similar when comparing election periods over time. Corresponding policy spaces, regression results, and methodological considerations are described in Appendix B.3. 54 Remember that standard errors barely change using seemingly unrelated regressions as the correlation between the residuals in the two dimensions is very low.

Results from selected RCVs
Before analyzing the effect of vote inequality on our aggregate measures of political orientation, it may be helpful to show that vote inequality affected voting behavior of MPs in specific, economically relevant roll calls. For this analysis, we selected five individual RCVs that could be unambiguously classified as liberal or conservative from four different legislative periods. Given the space limitations, we cannot provide a detailed summary of the parliamentary discussion surrounding each RCV. Note, however, that these roll calls are votes on important bills or amendments for which the divides in parliament became salient. Table 4 presents estimates from a variant of the baseline model in Equation 1 that regresses a binary vote outcome (yea/nay) from a single roll call on a Gini index of vote inequality and other constituency-level control variables. Results reflect a linear probability model where vote inequality affects the probability to vote 'yea'. Despite the potential classification errors described in Section 5 above, we determine whether 'yea' is a conservative or liberal position, as indicated in the lower part of the table, by the fact that the majority of MPs from conservative or liberal parties voted for the position. Across all 5 roll calls, we find that MPs from districts with a higher vote inequality have a significantly higher probability of voting for the liberal position.
Column 1 shows results from a RCV about the charging of tuition fees. Article 25, sentence 3 of the Prussian constitution stipulated that public primary schooling was free of tuition fees. Since 1850 municipalities were required to finance schooling from local sources or taxation. In 1869, the King proposed a bill to revoke this sentence. The RCV was a vote on the King's bill.
Column 2 shows results from a RCV within the debate about a major administrative reform (Kreisordnung). The amendment to the bill, introduced by a liberal MP, proposed that the newly introduced role of a public official (Amtshauptmann) should be filled following a selection by the local administration instead of an appointment by the King. It is essentially a vote about centralization vs decentralization of power.
Column 3 shows results from a RCV about the nationalization of railroads. It is essentially a vote on section 1 of the bill introduced by the Trade Ministry that determined whether the state should buy railroad companies.
Column 4 shows results from a RCV in the context of the debate about the famous income tax reform of 1891 (see Hill, 1892). MPs voted on an amendment to section 16 of the bill regarding the taxation of income from investment in joint stock companies (e.g. dividends). This conservative amendment requested a 3.5 percent deduction whereas a competing liberal amendment requested a 4 percent deduction.
Column 5 shows results from a RCV on a proposal to construct the Dortmund-Rhine segment of a canal connecting the rivers Rhine and Elbe, introduced by the Kaiser.
Conservatives opposed the canal construction because it would be detrimental to agrarian interest and would foster the import of cheap grain from the United States.
The findings allow us to discuss a thought experiment, in which we can understand how policies would have changed, if vote inequality had been different (ceteris paribus).
The results imply that when increasing vote inequality by one standard deviation (i.e., 0.039 Gini points), the probability of voting for the liberal position increases by 7 to 19 percent.
In two out of five RCVs such a change in inequality would have tipped the balance towards adopting the more liberal policy, in two cases the liberal majority would have been even larger, and in one case the conservative majority would have been substantially reduced.

Results when pooling RCVs across all election periods
Tables 5 and 6 present results from two independent OLS regressions as given in Equation 1, using as dependent variables the ideal points that reflect the two dimensions.

The liberal-conservative dimension
Column 1 in Table 5 presents the most parsimonious model explaining liberalconservative political orientation with two measures of inequality -vote inequality and landownership inequality. We find MPs from constituencies with a more unequal distribution of voters across classes to vote significantly more for liberal policies. In line with previous work, we also find MPs from constituencies with a more unequal distribution of landownership to have a more conservative orientation.
Subsequent columns add control variables that potentially affect the political orientation of MPs. By including them, we try to rule out that the effect of vote inequality is driven by economic development. Furthermore, we include several measures of cultural heterogeneity that are expected to correspond to secular-religious voting much more than liberal-conservative voting. 55 Columns 2 and 3 add the industrial employment share and the urbanization share to capture structural differences in the economic development of regions. More industrialized and urbanized constituencies are expected to be more progressive and thus prefer more liberal politics. Both variables are negatively associated with political orientation -55 While cultural factors such as religion and linguistic fractionalization are slow moving indicators that change only in the long run, development indicators such as urbanization and the size of the industrial sector can adjust more quickly and may therefore be endogenous to the political process. That being said, we are aware of the fact that some of our control variables are measured concurrent to political orientation and may be considered 'bad controls.' To address this issue, in Table D.1 in the Appendix, we show that results also hold in the smaller sample of Prussia in its pre-1867 borders, where we can use control variables from the censuses in 1849 and 1861, which pre-date our outcome variables.
i.e. predict that the local MP indeed supports more liberal policies. The coefficient on vote inequality is reduced but remains highly significant.
Columns 4-6 sequentially add the share of Protestants, the linguistic fractionalization, and the share of non-migrants to the model. Religious denomination and language heterogeneity pick up differences between the political orientation of Catholic regions in the West, the central Protestant regions, and the Slavic regions in the East. Since Protestant regions are more developed than Catholic regions, conditional on industrialization and urbanization, this variable may predict a more liberal orientation. Linguistic fractionalization is highest in the Slavic regions of Prussia that may oppose the national-conservative policies of Bismarck. A higher share of the population that lives in their municipality of birth is expected to reflect lower levels of openness and a more conservative orientation. However, these factors do not seem to play an important role in affecting liberal-conservative voting once a variable accounting for the literacy rate is included in column 7. The literacy rate is a significant predictor of conservative voting. This finding may be explained by the fact that conservative parties appealed to a more literate Protestant electorate. The literacy seems to capture conservative Protestants better than the overall share of Protestants which becomes significantly negative upon joint inclusion. 56 Across all specifications, we find vote inequality to be systematically associated with more liberal voting. In terms of magnitude, the coefficient on vote inequality in column 7 means that a one standard deviation increase in the Gini index is associated with a 0.18 standard deviation decrease in conservative orientation. 57 In other words, moving an MP from a constituency at the 10th percentile (∼ 0.47) of the vote inequality distribution to the 90th percentile (∼ 0.58), he will shift his orientation by 19.7% toward the liberal maximum.

The secular-religious dimension
In Table 6, we find that the relationship between vote inequality and secular-religious orientation is entirely spurious. While vote inequality predicts more secular voting behavior in initial columns, the coefficient is close to zero, once the share of Protestants of the constituency is accounted for. Not surprisingly, the share of Protestants is itself a strong predictor of secular voting behavior. This finding reflects the strong cleavage between Centre party MPs elected in the Catholic regions and MPs from Protestant regions that strongly favored a separation of church and state.
Furthermore, higher linguistic fractionalization is associated with more secular voting. In this dimension, linguistically more heterogenous regions encompass a larger share 56 This relates to the reasoning by Boppart et al. (2013) who find that differences in educational outcomes between Protestants and Catholics in late 19th century Switzerland are most pronounced in conservative milieus. 57 Vote inequality has a standard deviation of 0.039, liberal-conservative political orientation has a standard deviation of 0.395. Using the coefficient reported in column 7, the magnitude can be calculated as of minorities that may indeed favor policies that restrict a church that caters only to one religious denomination. Given the fact that more heterogenous regions usually host a predominantly Catholic Slavic minority which suffered from Bismarck's Germanization policies during the Kulturkampf, this is a surprising finding. In sum, this placebo test finds that the secular-religious orientation, which arguably transcended economic heterogeneity, is unrelated to vote inequality.

Robustness tests
This section tests the robustness of our findings against potential alternative explanations and confounding factors. Robustness tests are presented in Table 7, where Panel A displays results for the liberal-conservative orientation and Panel B displays results for the secularreligious orientation. For the sake of clarity in presentation, this table does not show coefficients of the control variables. We will mention them when needed below and will particularly focus on Panel A.
Column 1 repeats the baseline specification (column 7 from Table 5) before adding a control for per capita income tax payments to the model in column 2. By doing so, we address the concern that vote inequality is a mere proxy for the average income in an electoral constituency. 58 Results on vote inequality remain qualitatively unchanged but coefficients are slightly larger.
Column 3 adds a control for social uprisings that took place in a constituency be- After focussing on the local conditions of the constituencies, we would like to understand whether the results may be driven by observable characteristics of the individual MPs. If the peculiarities of the franchise system selected a certain type of MP into office, 58 Unfortunately, data on income tax payments are unavailable for constituencies that consisted only of one city (Stadtkreise = urban counties), which is why we lose these observations. The finding is robust to using a pre-1866 measure of income tax per capita which is observed for all constituencies, but not for the provinces annexed after 1866. 59 This finding is unaltered by using protests from the period 1867 to 1903 as the control variable.
this would constitute a mechanism through which vote inequality might affect political orientation. However, if MP characteristics reflect characteristics (and preferences) of the local elite, these may determine both inequality and political orientation and present a threat to identification.
Occupations in public administration and agriculture are associated with a more conservative orientation. Similarly, noble status and Catholic denomination are strong predictors of conservative voting behavior that explain a large share of the variation as measured by the R-squared. Interestingly, one of the few variables that predict more liberal voting is the dummy for MPs who do not reside in their own constituency. While electoral turnout is not related to voting behavior, a higher margin of victory predicts conservative orientation. MPs that participated in more roll calls vote more liberally, likely reflecting a more active role of liberal MPs in parliament, challenging the status quo. In sum, controlling for individual-level characteristics, the coefficients on vote inequality and landownership inequality are substantially reduced but the results remain qualitatively unchanged.
Regression results displayed in column 5 show that adding 35 administrative-district (Regierungsbezirk) fixed effects leaves the relationship of interest barely affected. Thus, broad regional disparities between the agricultural east and the industrial west do not explain the relationship between vote inequality and political orientation.
Above, we have shown that party affiliation is a strong predictor of voting behavior.
Our subsequent empirical framework has exploited variation across parties ignoring affiliation. Indeed, we think that party affiliation reflects much of the differences in political orientation that we are actually interested in. Controlling for party affiliation thus only allows us to exploit within party variation in political orientation which, by definition, is much more limited. Nevertheless, column 6 shows that, even conditional on party affiliation, MPs vote significantly more for liberal policies when elected in a constituency with a higher vote inequality.
The earliest available measure of vote inequality for all constituencies of post-1866 Prussia is from 1893. The data was collected just after the 1891 tax reform changed the distribution of voters across classes in the three-class franchise. We test robustness of the results using a Gini index of vote inequality from 1861 for a subset of constituencies reflecting the contemporary Prussian borders. Column 7 re-estimates the model with individuallevel controls in the smaller sample of 1861 constituencies. Column 8 shows that results are barely different from the findings in column 7 when using the 1861 measure of vote inequality.
Columns 1-8 in Panel B repeat the above regressions for the secular-religious (placebo) dimension. Well in line with our interpretation of this dimension, constituency characteristics capturing cultural factors (share Protestants, linguistic fractionalization and literacy rates) affect voting across all specifications. Among the individual-level controls in column 4, the only occupation significantly associated with religious policies is landownership, the only MP-level denomination that predicts religious voting is Catholic. MPs holding a noble title vote more pro-religiously, MPs holding an academic title vote more secularly. However, this does not change the relationship between vote inequality and secular-religious orientation. Similarly, adding administrative-district fixed effects, party affiliation, or replacing vote inequality with its 1861 measure does not affect the qualitative findings.

Mitigating potential endogeneity concerns
The Across Prussia, on average, the threshold for a first class vote was a payment of 56 thalers in taxes. However, regional variation was extreme: the minimum threshold was 7 and the maximum threshold was 12,496 thalers. If an individual tax payer contributed 60 Based on the classified income tax of 3 percent at the time, their income must have been 9,000 thalers. 61 The source does not clarify how one of the three voters with identical tax contribution of 270 thalers was selected to go into the second class, as ties were not regulated in the electoral law. more than half of the total local tax base, the second class even remained unpopulated.
Given that the local distribution of incomes was decisive, variation in residential location of high-income earners led to huge heterogeneity, especially in urban areas. Before a reform in 1891, municipalities first determined tax payment thresholds based on the city-wide population of voters, before subsequently breaking the electorate into wards. In wards with many high income voters, the first class was crowded, leading to a relative devaluation of a vote. In poor wards, the first and sometimes even the second class remained empty if none of the voters crossed the city-wide threshold. In many cases, but not in all cities, if a class remained unpopulated, voters were reclassified according to the ward level tax base, leading to a decrease of the minimum tax payments required to be allocated to the first class.
The above examples imply that there is strong heterogeneity in the distribution of voters across classes and that the allocation into the first class is beyond the power of the individual. In our standard regression framework, we aim at excluding systematic heterogeneity by including a substantial set of control variables to reduce endogeneity concerns.
However, to further address concerns regarding the endogeneity of vote inequality, we exploit variation in the share of voters in the first class in an instrumental variable approach.
We do so because the exact number of voters in the first class is arguably exogenous to local conditions. The argument here is not that such an approach is similar to randomly assigning votes to individuals but rather similar to arbitrarily changing voting power within the economic elite.
The exclusion restriction requires that the share of voters in the first class affects the political orientation of MPs only indirectly, through vote inequality. We argue that this assumption is not violated because it seems plausible that the share of individuals endowed with higher voting power because of the local tax distribution can only affect MP voting patterns via the inherent features of the three-class franchise. 62 Consider a hypothetical example of two constituencies with an identical number of tax payers. In both constituency A and B, the top two tax payers pay 200 and 100 thalers of tax, respectively.
Due to the distribution of tax payments of all other tax payers outside the top 2, in constituency A, the top tax payer (paying 200 thalers) is the only member of the first class, whereas in constituency B, the top two tax payers are both members of the second class. In other words, the share of tax payers in the first class is twice as high in constituency B than in constituency A. The exclusion restriction amounts to saying that the one and only firstclass voter in constituency A affects political outcomes in his constituency only by virtue of the fact that the resulting vote inequality inherent in the three-class franchise gives him more power, but not because his position in the income distribution per se has an effect on 62 Plausibility can be further improved by conditioning on our set of control variables.

RCV outcomes. Since the distribution of top incomes is identical in constituencies A and
B, this strikes us as a reasonable assumption. Table 8 presents results comparing OLS results to a 2SLS approach using the share of voters in the first class as an instrument for vote inequality in 1861. 63 We find the coefficient on vote inequality to be marginally smaller when comparing our baseline findings on the liberal-conservative orientation in column 1 to the findings in column 3. The qualitative findings however remain unchanged, even when controlling for individual MP characteristics and party affiliation. The results for the secular-religious orientation are close to zero in the 2SLS regression, again confirming the earlier findings of a lack of a relationship with vote inequality.

Discussing the mechanism
In this section, we explore our hypothesis (see Section 3) that industrialization created a level of income inequality that allowed large-scale industrialists to elect MPs who provided liberal policies that were conducive to industrialists. Therefore, we expect the effect of vote inequality on liberal voting to be amplified in areas with a higher share of large firms.
For this exercise, we digitized data on the number of large and small firms in a Prussian county, from the 1875 Census of Industrial Firms. In the census, large firms are defined as firms with at least 5 employees.
To probe the hypothesis, we define a dummy for the upper quartile of the distribution of the share of large firms. 64 Table 9 shows the result from this exercise. In column 1, we add the main effect of being in the upper quartile in terms of share of large firms as a regressor. The variable by itself does not affect voting behavior. In column 2, we estimate an interaction term between vote inequality and the large firms dummy. We find that vote inequality has a stronger effect on liberal voting in regions with a higher share of large firms, confirming our hypothesis. 65 Note that our set of controls includes the industrial employment share, thus holding the general level of industrialization fixed. Finally, columns 3 and 4 show that the concentration of large-scale industry does not matter for voting along the secular-religious placebo dimension.
We interpret these findings as evidence that voters were better able to coordinate on liberal policies when the electoral elite was small and uniformly benefited from a liberal economy. While we do not observe the characteristics of the electoral elite in our data, we think it is conceivable that in some regions industrialization created a level of income inequality that allowed large-scale industrialists to use the political power provided by the 63 Results are qualitatively similar when using 1893 vote inequality. Results are also similar when using the share of voters in the first and second class as an instrument. 64 Alternatively, we can base quartiles on the number of large firms in a county and find very similar results. 65 Table D.4 shows similar results across election periods. franchise system to commission a self-serving liberal agenda.

Results comparing election periods
In this section, we estimate models based on Equation 1 using the political orientation of MPs estimated separately from RCVs for each election period. By doing so, we can inspect the stability of the relationship between political orientation and vote inequality over time.
Given changes in party composition and coalitions in Parliament during the 1870s, a stable relationship would be corroborating evidence of vote inequality as an important (and stable) driver of political ideology. Indeed, results when comparing the relationship across election periods broadly confirm the patterns established in the pooled sample. However, due to lower number of RCVs that take place in a single election period compared to the pooled period, estimates of ideal points are less precise and have much more measurement error. Most likely this explains the variation of coefficient magnitudes across election periods and the larger standard errors compared to the model that pools all RCVs.
Results showing the relationship between liberal-conservative orientation and vote inequality, conditional on constituency characteristics, are presented in Table 10. We subdivide election periods between a Kulturkampf -period, a Protectionism-period, and a post-Bismarck-period, reflecting political life of Bismarck and our discussion of key debates in Appendix B.2. Across election periods, we find the same indicators to consistently predict liberal-conservative voting behavior. For all election periods, we confirm earlier findings that higher vote inequality predicts more liberal voting behavior. Similarly, we find landownership inequality to be positively associated with conservative voting behavior throughout all election periods. Furthermore, MPs in constituencies with higher shares of industrial workers and urban dwellers lean towards liberal voting.
Results for the secular-religious orientation, are presented in Table 11. We find that vote inequality is significantly related to pro-religious orientation in two periods (10 and 14). In all other periods the relationship remains statistically indistinguishable from zero. Interestingly, we find landownership inequality to show the opposite sign to vote inequality and to switch signs correspondingly. The only variable that significantly predicts voting behavior in this dimension throughout all election periods is the share of Protestants.
Linguistic fractionalization and industrial employment show a stable negatively association with support for the church, while there is no stable pattern for other variables.
All findings are confirmed when controlling for individual MP characteristics in Tables D.2 and D.3 in the Appendix.

Conclusion
This paper challenges the predominant portrayal that the inherent inequality of the Prussian three class franchise system produced a conservative parliament. Exploiting the voting behavior of MPs in 329 roll call votes to predict political orientation using the optimal classification method Poole (2000), we find that higher vote inequality across the three franchise classes is associated with a more liberal orientation. This finding contrasts with the historical narrative arguing that the system of franchise produced more conservative votes by providing the landed elites with all the political power. Additional evidence shows that the results are driven by constituencies with a high concentration of large industrial firms, corroborating our interpretation that large-scale industrialists were able to take advantage of the franchise system and elected delegates who voted for liberal policies. This finding also lends support to the view that historically, economic elites were not a monolithic group and some may have favored liberalization and modernization for self-serving reasons.
Furthermore, this paper is the first to present a comprehensive picture of the Prussian policy space over a period of ten election periods , covering a crucial period in the structural transformation of the German Empire from an agricultural to an industrial economy. Despite the finding that vote inequality was not responsible for a conservative orientation of MPs, the stability of the policy space over time is consistent with the idea that the system produced a parliament that was able to avoid substantial changes to the franchise until WWI.

Election periods 10−19 (1867−03)
First Dimension (Liberal − Conservative) Second Dimension (Secular − Religious)   Notes: This table reports the party affiliation of all 3658 MP-by-period observations whose votes were scalable. Column Party depicts the original party assignment according to Kühne (1994b). Cases in which party affiliation was uncertain were coded according to the most likely party affiliation. Column Classification consolidates smaller parties with similar platforms. Column Indicator lists a party code used in graphical presentation of the policy space. Columns Freq. and in % show the frequency and share of party affiliation in the pooled sample. MPs with a minimum requirement of 15 votes. *(Due to the low number of RCVs in period 13, the minimum required votes was reduced to 12 to allow scalability of MPs.) Columns 5 and 6 list the percentage of correctly classified yea and nay votes and the Average Proportion Reductions in Error (APRE) predicted by the optimal classification process for the one-dimensional case. Columns 7 and 8 list the same for the two-dimensional case. Columns 9-11 list the eigenvalues for the first three dimensions. The bottom row show results when using the pooled sample.        1 and 6 repeat baseline results in the 1861 sample. Columns 2-3 and 7-8 report the first and second stages of a 2SLS approach using the share of voters in the first class as an instrument for vote inequality. Columns 4 and 9 add the full set of MP controls. Columns 5 and 10 add a set of party affiliation controls, defined as the share of active election periods a MP was affiliated with each of the parties. Standard errors, clustered at the constituency level, in parentheses. *** denotes statistical significance at the 1% level, ** at the 5% level, and * at the 10% level. ) is equal to one if the constituency-level share of firms with more than 5 employees is in the highest quartile. Standard errors, clustered at the constituency level, in parentheses. *** denotes statistical significance at the 1% level, ** at the 5% level, and * at the 10% level.

A.1 Constituency-level variables
The following variables were either part of the Galloway database (see Galloway, 2007), the iPEHD database (see Becker et al., 2014) or specifically digitized for this paper. The corresponding original sources are indicated below.
Vote inequality. Gini coefficient, calculated using the number of voters in each class in the three-class franchise system of 1893 assuming that the tax burden of each class amounts to exactly one third, constructed using county-level data reported by Königlich Preußisches Statistisches Bureau (1864-1905, vol. 17).
Vote inequality 1861. Gini coefficient, calculated using the number of voters in each class in the three-class franchise system of 1861 assuming that the tax burden of each class amounts to exactly one third, constructed using county-level data reported by Königlich Preußisches Statistisches Bureau (1861-1904, vol. 2).
Landownership inequality Gini coefficient, calculated using using the number of farms in 6 size groups in 1882. Farms with arable land up to 1 hectare, 1 to 2 ha, 2 to 10 ha, 10 to 50 ha, 50 to 100 ha, and more than 100 ha. Constructed using county-level data reported by the Königlich Preußisches Statistisches Bureau (1861-1934, vol. 76c).

Linguistic frac. (Herfindahl).
The Herfindahl index of (ethno-)linguistic diversity calculated using the number of a county c's inhabitants whose mother tongue was German or any of the k = 20 other languages p. Constructed using the following equation

A.2 MP-level variables
The following variables were taken from a TUSTEP dataset kindly share by Thomas Kühne.
Publ. admin. Unity (by-period analysis) / the share of election periods (pooled analysis), if a MP reported an occupation in the public administration, public service, public judiciary, or military.
Agriculture. Unity (by-period analysis) / the share of election periods (pooled analysis), if a MP reported to landownership or peasantry as main occupation.
Industry. Unity (by-period analysis) / the share of election periods (pooled analysis), if a MP reported to be occupied in manufacturing or industrial production, to be a merchant, banker, or shipmanager.

Service. Unity (by-period analysis) / the share of election periods (pooled analysis), if a
Elec. turnout. The share of electoral delegates that voted in the election for the members of the Prussian House of Representatives, divided by the total number of electoral delegates (in the pooled analysis, we use the mean turnout across election periods).
No. of MPs. The total number of MPs elected to the Prussian House of Representatives from the constituency.    Drawing on biographical information from Kühne (1994b), we report the party affiliation of MPs in Table 1 In some cases when party affiliation was uncertain, we used the most likely party affiliation given in Kühne (1994b). In cases of changes in party affiliation within a legislature, we use the initial party affiliation. Party changes within legislature are extremely rare and affect well below 1% of MPs, on average. In the pooled analysis, in cases when MPs changed party affiliations between legislatures, we assigned the party for which the most roll call votes were cast. 67 More than 30% of Prussian MPs were members of both parliaments at some point in time. Concerns that MPs who belonged to both parliaments were systematically different can be addressed with a dummy variable for dual membership. In unreported regressions, we find that adding this control variable does not change our results. 68 In that sense, the Prussian Parliament as a whole was a conservative one. But as argued before, our interest is in the link between vote inequality and political orientation within the three-class franchise.
(R) and had shifting alliances with the National Liberal Party (N) and the Centre Party (Z).
The Conservative Party was conservative in all dimensions, aimed at retaining the feudal structure, the monarchy, and opposed the unification of the German Empire in favor of a strong Prussian state. Conservatives were strongly in favor of agrarian protectionism, and weakly in favor of industrial protection and guilds. They were divided over the Kulturkampf in the 1870s, but rallied behind Bismarck after his 'conservative turn' in 1879, and supported the return to protectionism in the 1880s. The Free Conservatives (R), con- This description shows that liberal parties of the time did by no means stand for extreme economic liberal views as one might find in some Western Democracies of the 20th and 21st century. Liberal parties were more liberal in comparison to a conservative spectrum that aimed at retaining the feudal structure, the monarchy, and even opposed the unification of the German Empire. It is important to keep this in mind when we talk about liberal-conservative policies and MPs.

B.2 Key debates in the period 1867-1903
Our empirical analysis starts in 1867, after Prussia won the Austro-Prussian War of 1866, and annexed the Kingdoms of Hanover, the Electorate of Hesse, and the Duchy of Nassau to reach its maximum extension which it kept until WW1. From 1867, i.e. election period 10, the new provinces of Hanover, Hesse-Nassau and Schleswig-Holstein were part of Prussia and elected members to the Prussian House of Representatives. Election periods initially lasted three years and, after 1888, five years. Table B.1 shows the dates of primary elections (column 1), of the election of MPs by the electoral delegates (column 2), and the start and end dates of the legislatures (columns 3 and 4). Column 5 lists the number of RCVs that took place in each election period. We end our analysis in 1903 due to a range of changes in the electoral system and the division of constituencies thereafter. Table B.2 presents a systematic overview over the content of all RCVs comparing the ten election periods. 69 A fundamental issue looming throughout the entire period under analysis is the social conflict arising between advocates of economic liberalism and advocates of a social policies. This conflict did not arise in the Prussian parliament due to the lack of representation of socialist parties and was eased by the introduction of Bismarck's system of social insurances from the 1880s. were not yet fully developed and the preceding election periods were dominated by a constitutional conflict between the Emperor and Parliament which had power over 69 The categories in this table roughly correspond to the ministries of the Prussian government. Due to the small number of votes on foreign affairs (mostly a Reichstag matter), this category was combined with interior affairs. Due to the high frequency of budgetary issues and procedural issues, these items form a separate category. RCVs were first itemized by keywords according to the original parliamentary minutes before keywords were assigned to a category. the budget. The conflict, although officially resolved in 1866, continued in election period 10, which stands out as a period mainly concerned with budgetary issues (budgetary allocation to different ministries).
The following three election periods 11 to 13, covering the years 1870 to 1879, are dominated by the struggle between the Prussian state and the Roman Catholic Church (ca. 1871-1878, and finally diplomatically settled in 1887). During this period, labeled as Kulturkampf, a large share of RCVs are directly related to debates about the so called May Laws (Maigesetze) designed to bring education, marriage, etc. under the control of the state. We will see that this has implications for our estimation of political orientation by ideal points. The debates about the separation of church and state do not follow a classical liberal-conservative divide, but define a separate political dimension relevant to explaining voting patterns. Apart from the Kulturkampf, a range of RCVs can be found in the category Interior and Foreign that are concerned with administrative reforms of the counties. The county reform was heatedly debated as it also determined the fiscal responsibilities for welfare benefits.
We use the label Protectionism for a second subperiod to characterize the three election periods 14 to 16 covering the years 1879-1888. Following his so-called 'conservative turn' in 1879, Bismarck broke with the National Liberal Party and tended towards more nationalist-conservative policies. It is important to note that, given the subsidiarity of the Prussian House of Representatives to the German Reichstag, some policy issues, especially those related to external affairs, were debated primarily in the Reichstag. For instance, the turn from liberalism to protectionism and the "Marriage of Iron and Rye" (see Schonhardt-Bailey, 1998;Lehmann, 2010b) were issues mainly debated in the Reichstag. 70 However, nationalization of railways, a Prussian matter, has a similar flavor in terms of beliefs in free markets vs state, and is the topic of a substantial number of RCVs in period 14 (see category Commerce). Similarly, resonating with the findings by Cinnirella and Schueler (2018), some RCVs were concerned with the introduction of German as main language of instruction in regions with Slavic majority (see category Culture in election period 15) and the promotion of German interests and settlement in eastern Prussia (see category Interior and Foreign in election period 16). Although the Kulturkampf was largely settled, we find many RCVs related to church and state in election periods 14 to 16. These are predominantly concerned with mitigating the impact of the May Laws, e.g. pardoning of wrongdoing according to these laws and sorting out remaining issues related to education.
Finally, we use the label post-Bismarck for a third subperiod that combines the final three election periods 17 to 19 (1888-1903). With the elections in 1888, legislatures move 70 Agricultural tariffs were the matter of four RCVs in the Prussian House of Representatives. One in election period 15, one in 16, and two in 19, but -in line with subsidiarity -voting was concerned with submission to the Reichstag.

B.3 The policy space -comparing across election periods
While not at the core of our research question, we can compare regression results over time. Such an analysis may however be complicated if topics debated in parliament change considerably across election periods such that some periods might be dominated by the Kulturkampf whereas other periods are more dominated by economic issues. Since the optimal classification algorithm is agnostic about the content of debates and corresponding RCVs, the two dimensions may be rotating. The first dimension will always load on the dominant cleavage between political parties in each individual election period. Thus, we expect to find switches in the ranking of dimensions across periods when the dominance of certain topics changes and parliamentary front-lines shift. A regression analysis comparing individual election periods however rests on the assumption that the dimensions can be interpreted in the same way across periods. We resolve this issue by rotating each individual policy space to fix the centroid of the Conservative Party to the horizontal axis to generate the ideal points for the first dimension. We then rotate the policy space to fix the centroid of the Centre Party to the vertical axis to generate ideal points for the second dimension.
Further comparability issues will arise if the content of RCVs shifts to topics that do not relate to the two dimensions of interest. This cannot be judged from inspecting the frequency of RCVs by topic discussed in Section B.2, as even seemingly unrelated issues can be dominated by the two dimensions. 71 Since we cannot ensure that the two dimensions load on the same content of debates over time, we resort to inspecting the graphic representation of the policy spaces, and to regression results using party dummies. 71 For instance, an RCV on hunting rights might, at first sight, seem unrelated to both liberal-conservative as well as secular-religious cleavages. In reality, roll call voting might still be dominated by those two dimensions. to 1903 rotated to fix the centroid of the Conservative Party to the horizontal axis. Comparing the plotted policy spaces, we find the same patterns as in the pooled sample. Fellow party MPs cluster along party lines on the one hand, but also display significant withinparty heterogeneity in both dimensions in almost all election periods. With the exception of period 13, when only 15 roll calls were cast, we find election periods to provide enough variation in voting behavior to even allow for a period-specific econometric analysis below.
The plots show a clear divide between liberal and conservative parties on the horizontal axis and a clear divide between secular and religious parties on the vertical axis. 72 The distribution of parties within the policy space confirms the assumption that the two dimensions show considerable stability over time. 73 We test the accuracy of the rotation for the purpose of our research question, by regressing ideal points capturing the first dimension on party dummies in Table B Similarly, Table B.4 presents results from regressions of ideal points capturing the secular-religious dimension on party dummies. The pattern of results depicts a sorting of parties along this dimension. When comparing across election periods, we find a considerable wedge between the Centre Party and most other parties arising from election period 72 Although we rely on the exact same MPs as anchors across the majority of periods, when inspecting the pre-rotation policy spaces in Figure D.5 in Appendix D, we indeed find that the first dimension and second dimensions are switching over time. In some election periods, the first dimension loads more heavily on secular-religious debates whereas the second dimension loads on liberal-conservative debates and vice versa. Furthermore, in many periods, RCVs seem to load on both dimensions and parties seem to align with a diagonal orientation. 73

Election period 16 (1886−88)
First Dimension (Liberal − Conservative) Second Dimension (Secular − Religious) Positions of MPs in the Prussian House of Representatives. Each shape represents the political orientation of an MPs during one of the election periods according to voting behavior in roll calls during that period. Each shape is rotated to fix the centroid of the Conservative Party (K) to the horizontal axis.  Notes: This table reports the content and frequency of roll call votes by category and election period. Categories roughly correspond to the ministries of the Prussian government. Due to the rare occasion of votes on foreign affairs, this category was combined with interior affairs. Due to the high frequency of budgetary issues and statutory issues, these items received a separate category. RCVs were first itemized by keywords according to the original parliamentary minutes before keywords were assigned to a category.     Notes: This table reports results of OLS regressions of secular-religious orientation on party dummies by election period. Positive values of the dependent variable are interpreted to reflect higher levels of adherence to the church; negative values reflect secular orientation. Omitted Party = Zentrum (Z). Standard errors, clustered at the constituency level, in parentheses. *** denotes statistical significance at the 1% level, ** at the 5% level, and * at the 10% level.

C Voting behavior in individual roll calls
To illustrate how we operationalize political orientation using the two-dimensional policy space, we select two specific RCVs, one on financing of poor relief (vote 6 in the election period 11, 1870-73) and one on church autonomy under state supervision (vote 29 in the same election period). Figure                      Positions of MPs in the Prussian House of Representatives. Each shape represents the political orientation of an MPs during one of the election periods according to voting behavior in roll calls during that period. Shapes are not rotated.   Positions of MPs in the Prussian House of Representatives. Each shape represents the political orientation of an MPs during one of the election periods according to voting behavior in roll calls during that period. Each shape is rotated to fix the centroid of the Centre Party (Z) to the vertical axis.    ) is equal to one if the constituency-level share of firms with more than 5 employees is in the highest quartile. Standard errors, clustered at the constituency level, in parentheses. *** denotes statistical significance at the 1% level, ** at the 5% level, and * at the 10% level.