From Spatial Distance to Programmatic Overlap: Elaboration and Application of an Improved Party Policy Measure

From Spatial Distance to Programmatic Overlap: Elaboration and Application of an Improved Party Policy Measure Martin Mölder June 6, 2013 Abstract In contemporary representative democracies the political profiles of parties have become central to our understanding of party systems and governing. Although there are various approaches for the measurement and interpretation of party policy, there is currently no available measure that would be both valid and applicable for an extensive crosscountry time-series analysis. Expert surveys lack the necessary scope in time and the left-right index (RILE) of the manifesto dataset lacks empirical validity. The current project will argue that the prime reason for this is the use of the spatial metaphor for party policy interpretation. The project will thus consist of two inseparable objectives. First, to propose an alternative measure of party policy, which can be constructed on the basis of currently available data, but would avoid the issues of applicability and validity of current measures. Second, to demonstrate how this measure can be applied in party system polarization and coalition formation research, comparing it to existing policy measures. 1

Contents 1 Prelude: Problems of Scope and Validity in Party Policy Analysis 3 2 Overview of the Project 4 3 The Analysis of Party Policy Theoretical and Empirical Background 5 3.1 Parties, Policy and Representative Democracy......... 5 3.2 The Spatial Metaphor of Party Policy............. 6 3.3 Current Measures of Party Policy................ 8 3.4 Applications of Party Policy Measures............. 11 3.5 Problems of Measuring Difference through Location..... 12 4 Proposal for a Measure of Programmatic Overlap 13 4.1 A Measure of Programmatic Overlap Based on Manifesto Data 14 4.2 Future Perspectives Computerised Measures of Overlap.. 15 4.3 Limitations of the Proposed Measure.............. 16 5 Applications I: New Perspectives for Characterising Party Systems 16 5.1 Measuring Overlap Instead of Polarization........... 16 5.2 Measuring Party System Policy Change............ 17 6 Applications II: Understanding Coalition Formation 18 6.1 Coalition Formation and Programmatic Overlap....... 18 6.2 Explaining the Political Homogeneity of Government Coalitions 19 7 Plan of the Project 20 References 21 2

... the usefulness of models depends absolutely on the interchange between theory-building and empirical observation. (Stokes 1966, p. 178) 1 Prelude: Problems of Scope and Validity in Party Policy Analysis If a social scientist would currently want to analyse party policy covering both East- and West-Europe over an extended period of time, let alone think of including countries beyond this region, there would not be a single valid measure that would be available for use. Of the two most common measures, expert surveys (like Benoit and Laver 2006; Bakker et al. 2012) have a limited scope in time and cannot be done retrospectively and the left-right index (RILE) of the manifesto dataset (Budge and Klingemann 2001; Budge et al. 2001; Klingemann et al. 2006) is problematic when applied across Eastern and Western Europe (Savage 2012) and has issues with adequately positioning 1 parties (Klingemann et al. 2006, Chapter 4; Dinas and Gemenis 2010; Gemenis and Dinas 2010). Computerized content analysis of political documents (Laver, Benoit, and Garry 2003; Slapin and Proksch 2008), which can promise both wider scope and greater reliability, has issues of its own that have prevented its wider use. Therefore, any kind of analysis that would want to apply party policy as a variable and aim for a wider analysis would be faced with the options of either using invalid measures or forgoing the research. In this context, the aims of the current project are: First, to propose a measure of party policy, which can be calculated on the basis of available data and would avoid most of the validity issues of existing measures. Second, to demonstrate the application of such a measure in estimating party system polarization and explaining coalition formation. It will be argued below, that much of the problems of current measures derive from the spatial metaphor for interpreting party policy, which transfers properties of physical space to political space. Physical space is given and constant, policy spaces, however, are constructed. This has implications for the nature of the space as well as policy dimensions that are supposed to serve as points of reference for locating parties. 1 Here and below inverted commas will be used in the case of the spatial metaphor related to the interpretation of party policy. 3

2 Overview of the Project Party policy is central to the analysis of contemporary democratic politics. On the one hand, the policies of parties and the differences between them address some of the core theoretical issues of party behaviour (e.g. Strom 1990; Strom and Müller 1999) and the functioning of representative democracies (e.g. Dalton, Farrell, and McAllister 2011). On the other hand, reliable and valid data about party policy is essential to empirical analyses of party systems and democratic governance. It directly relates to research on party system polarization (Dalton 2008) and coalition formation (Martin and Stevenson 2001; Strom, Müller, and Bergman 2008; Andeweg 2011) among other areas. In this context, the rest of this paper constitutes an overview of the proposed project which can be divided into the following four parts. First an overview of current ways of policy measurement and interpretation will be given, starting from the changing role of party policy in our understandings of democracy. One the one hand, our ways of thinking about party policy are fundamentally shaped by the spatial metaphor, which aims to locate parties on policy dimensions. On the other hand, our needs for measures of policy derive form the role of policy in empirical research, which does not presume a spatial understanding of policy and is mainly just interested in how different parties are from each other. Thus, second, the project will propose a measure of programmatic overlap based on the election manifestos of parties, which would avoid the issues of validity and inapplicability of current measures. Third and fourth, the project will show how such a measure can be used for party system and coalition research. The applicability of this measure is broad and the plan of the project will address only a few of the possibilities. It must be acknowledged that the exact nature of the empirical applications is likely to evolve with the project. The most promising avenues will be included here. A measure of programmatic overlap can be used as an alternative to estimates of party system polarization. The proposed measure will be compared to existing estimates of polarization and tested in models that have used them in explaining voting behaviour. Programmatic overlap can also be used to calculate change in party systems from election to election, which would be a necessary and novel complement to indicators of electoral volatility. Finally, it will be shown how the measure of programmatic overlap can be used in models predicting coalition formation. Using exactly the same models and data, but different variables for political similarity/difference, it will be possible to evaluate the performance of the proposed measure. Furthermore, it will be shown how this alternative measure can be used to explain a particular kind of coalition government the so called non-connected coalition. 4

3 The Analysis of Party Policy Theoretical and Empirical Background Analysis of party policy has become a notable sub-area of political science. This has involved measuring the locations of political parties and applying them for characterising the political landscapes of party systems or explaining processes of governing. Having valid and applicable measures of policy is therefore not only a question of method and measurement, but also a problem that effects the quality of work across party and party system research. 3.1 Parties, Policy and Representative Democracy The role of party policy has changed considerably throughout the development of modern political science. In the middle of the twentieth century virtually all main authors left party policy to the sidelines. From roughly the 1970s onwards research on parties started to focus more and more on the policy of and policy relations between parties. By now party policy research has become a firm part of political science. E. E. Schattschneider s book Party Government (Schattschneider 2004) can be considered as the symbolic watershed, after which political parties were established at the core of the concept of democracy. Not long before him many significant authors on politics (Ostrogorski 1902; Michels 1915) had considered parties as an impediment to democracy. After that time competition between parties for power in the writings of classical authors such as Schumpeter (2006), Downs (1957), Epstein (1967) and virtually all who came afterwards was at the heart of the idea of democracy in a modern state. From the 1970s onwards, however, party policy in addition to just power was becoming more and more central. Sartori, in his classical work considered parties as being essential channels of expression (Sartori 2005, pp. 24-25). Ideological disagreements in a fragmented party system were for him of crucial importance for the functioning of democracy. Robertson (1976), writing also in the 1970s, noted that party competition was first about selecting governments and only secondarily about representation (ibid., p. 14). The mechanisms of democracy for him nevertheless relied on the hopes that policies desired by the voters would be adopted (ibid., p. 21). For von Beyme (Beyme 1985) party policy was at the centre of analysis and parties for him were primarily ideological organizations. He understood the main function of parties to be interest aggregation, which included the identification of party goals in the form of an ideology or programme (ibid., pp. 11-13). Thus party policy slowly moved to the centre of much of party system research. Many recent writings on concepts of representation and accountability (Urbinati and Warren 2008; Andeweg and Thomassen 2005; Lupia and Mc- Cubbins 2000; Strom, Müller, and Bergman 2003) set the role of policy in 5

party interaction, behaviour and voter-party relations at the centre of what democracy should be. Even if thinking in strictly procedural terms, the idea of responsiveness to voters (Saffon and Urbinati 2013, p. 462) is something that implies the flow of policy from voters through parties to policy output. Thus, although the pursuit of policy is only one possible goal that parties could have (Strom and Müller 1999; Reniu 2011, pp. 116-120), party policy is at the core of our normative-theoretical understanding of democracy by now. 3.2 The Spatial Metaphor of Party Policy We have come to think of party policy in spatial terms and this has fundamental implications for how it has been measured and interpreted. The spatial metaphor assumes that the domain of policy is a space akin to physical space with identifiable locations of parties that can be determined as coordinates on a set of dimensions and with distances between parties, which can be measured like continuous distances in physical space to indicate difference between parties. The spatial metaphor was introduced to political science by Anthony Downs (1957), who adopted the metaphor from the economists Hotelling and Smithies (ibid., p. 115). The original objective of Hotelling in formulating his theory of spatial competition between firms (Hotelling 1929) was the problem of stability in a duopoly. He used the example of sellers positioning themselves in physical space in order to attract the maximum number of customers. He drew a parallel to the American political system with two parties competing for votes in a political space. Smithies (1941) maintained and elaborated the mapping of such ideas of competition from physical space to political space. Neither Downs nor the economists considered the implications that this metaphor would entail for thinking about political space. This has also been true for many of the influential authors who adopted the metaphor. Thus, for Robertson thinking in terms of space and spatial dimensions was a natural and efficient way of dealing with a complex phenomenon and something that was basic to our conceptualisation of political competition (Robertson 1976, p. 55). Sani and Sartori in their study of party system polarization adopted the left-right dimension, the simplest version of the spatial metaphor 2, together with its implications automatically and without comment (Sani and Sartori 1985), despite the fact that Sartori in his earlier work had some reservations against the spatial metaphor (Sartori 2005, pp. 290-291). Also von Beyme considered the left-right dimension a valuable orientation aid (Beyme 1985, p. 136). In subsequent research, the spatial metaphor in the form of the left-right dimension, if not a multidimensional 2 In the case of a single dimension, the metaphor of spectrum has also sometimes been used (Sartori 2005, p. 111). 6

space, has become a standard way of thinking about policy, expanding from the domain of party system research also into voting behaviour (Enelow and Hinich 1984; Evans 2004). 3.2.1 The Problem of the Nature of Space The original idea of spatial competition in economics as developed by Hotelling (1929) and Smithies (1941) was not a metaphor. The space that they were talking about was actual physical space. If we talk about policy and ideology, however, space becomes metaphorical and it should be asked if it does justice to truth to transfer the qualities of physical space to political space. Physical space is the same everywhere. In contrast, political spaces can be highly variable (Stokes 1966, p. 168). This raises the question of whether we can think of these spaces as the same everywhere and for everyone and what exactly their specific content is (past perceptions, current saliency or policy preferences) (Daalder and Mair 1985, p. 20). Furthermore, physical space is measurable on a continuous scale, which has been transferred to political space. It has been noted, however, that this assumption is violated by the reality of policy alternatives, which tend to be categorical by their nature (Stokes 1966, p. 170). Therefore, the nature of political space is ambiguous and variable, raising problems of interpretation discussed below. 3.2.2 The Problem of the Number of Dimensions In physical space the number of dimensions is given, fixed and always the same. As long as travelling and transportation costs are concerned (main factors related to space in the economic example of Hotelling (1929)), no more than two dimensions and for most purposes no more than one, is necessary. Distance is a line that can be represented in one dimension. For policy space, things are more complicated. If we consider dimensions to be separate sources of variation of interest (Jacoby 1991, p. 27), then determining the number of dimensions can be rather problematic as it depends not only on the objective of the analyst, but also the nature of the specific space. Assuming one dimension is common, but empirical analyses have shown time and again that assuming a single dimension is not empirically grounded and that policy spaces tend to be rather complex and multidimensional (Daalder and Mair 1985; Stokes 1966; Albright 2010). At most we have to conclude that the dimensionality of policy spaces cannot be presumed and taken for granted as is the case with physical space. 7

3.2.3 The Problem of the Frame of Reference Finally, if we assume or conclude that a certain set of dimensions with a certain substantive content exists, we also have to argue for their empirical validity (Stokes 1966, p. 173). Who are the actors that perceive these dimensions as such? There can be remarkable divergence of perceptions between voters and politicians for there is nothing to guarantee that they have a common understanding. Furthermore, the analyst who determines or interprets the dimensions adds yet another frame of reference. It cannot be assumed that parties or voters have the same benchmarks as the analyst for determining the locations of parties. 3.3 Current Measures of Party Policy Increasing interest in party policy has generated numerous measures. The following will consider only the most widely used. For analytical purposes, it is important to distinguish between sources of data and the way this data is used to determine political positions. For the latter the differentiation by Benoit and Laver (2012) between measures that require a priori interpretation (deductive) and measures that rely on post hoc interpretation (inductive) is used. 3 3.3.1 Sources of Data Information about party policy has generally been obtained from three sources - expert (or mass) surveys, textual data or party behavioural data. Each has its advantages and limitations for constructing valid measures of party policy. Information obtained from expert surveys is perhaps one of the easiest and least costly sources of information (Benoit and Laver 2006, p. 75; Volkens 2007). Two of the most prominent expert surveys of party policy have been the Chapel Hill survey (Bakker et al. 2012) and the Benoit-Laver expert survey (Benoit and Laver 2006), a continuation of a study conducted in the beginning of the 1990s (Laver and Hunt 1992). An expert survey, which is both a source of data and a mode of interpretation at the same time, consists in contacting a number of country experts and asking them to place the parties on certain pre-determined scales. This can be an unspecified left-right dimension or more specific scales corresponding to policy domains (Benoit and Laver 2006, pp. 83-87). Either way, the frame of reference is determined together by the researcher and the expert and it can be assumed to be applicable at best only for the present and not validly extendible to the past (e.g. implicitly suggested by Volkens 2007, p. 113). Expert surveys 3 It should be noted that this distinction is ultimately a matter of degree and each has elements of both. 8

have developed a certain foundational status among policy measures as they are habitually used to cross-validate other measures of party policy (Benoit and Laver 2006, p. 75; Lowe et al. 2011). Also indicative of the political profile of a party can be the voting behaviour of its MPs. Thus, roll-call voting data for which it is known how each MP has voted is one possible source of data for party policy (Benoit and Laver 2006, pp. 69-71). This has especially been used to analyse the political positions of politicians in the US (ibid., p. 69). It has been noted, however, that this data does not give a valid picture in the case of multi-party systems with coalition governments where voting occurs along coalition lines (ibid., p. 70). Therefore, this source of data, like the expert survey, has limited applicability. One of the most prominent sources of data for party policy is textual data. Politicians and party leaders make political statements that can be considered indicative of the party position. Of all the possible textual sources, the most widely used for both human and computerised content analysis has been the party election manifesto. Party manifestos are regular statements of party policy that are attributable to the whole party (Budge and Farlie 1985, pp. 272-273; Robertson 1976, p. 72) and the only message that is in the full control of the party (ibid., p. 12). 3.3.2 Deductive Approaches to Positioning Two of the most prominent ways of positioning parties in policy spaces the RILE index of the manifesto dataset and various expert surveys are clearly deductive. Dimensions of policy are either explicitly or implicitly determined and interpreted before empirical positioning. As expert surveys are of limited application simply because of their scope as a source of data, only the RILE index, but also computerised content analysis, are considered below. The left-right index (RILE) is based on the manifesto dataset, which is the most extensive source of information on party policy. It is a content analysis (Budge and Bara 2001, pp. 4-5; Werner, Lacewell, and Volkens 2011), which brakes each election manifesto into quasi sentences and assigns each into one of 56 categories of policy issues. The index is constructed as follows (Budge and Klingemann 2001). A fixed definition of the opposite poles of the left-right dimension is assumed. Thereafter, two sets of 13 categories of the 56 in total are determined as representative of either pole. Finally, the proportion of left categories is subtracted from the proportion of right categories. The resulting value is supposed to represent the position of a party on the left-right dimension. In order to avoid some of the shortcomings of the index, a more intricate version of essentially the same logic has been suggested (Lowe et al. 2011). Recently the use of the manifesto data has been further developed to estimate the positions of parties 9

on certain pre-determined dimensions (Bakker 2009; König, Marbach, and Osnabrügge 2013; Elff 2013), but the methodological complexity and assumptions (at times highly questionable) of these positionings render their results largely uninterpretable. From the two most prominent computer content analysis approaches, one, Wordscores (Laver, Benoit, and Garry 2003), is distinctly deductive. It relies on a set of reference texts, for which a political position on a dimension of interest is predetermined. The program compares texts with unknown positions to the reference texts to produce a measure of position. Thus, although the process is automated, it still relies on human a priori interpretation, which determines the values of the analysed texts. One of the main advantages of all deductive approaches is that they do not require post-hoc interpretation and thus there is less ambiguity and guesswork (Benoit and Laver 2006, p. 76). The researcher assumes that certain dimensions with a definite substantive content exist. The structure and content of the political space is thus presumed and a priori given and what is left for empirical analysis is simply to fix location of parties on these dimensions. This, however, is also the biggest weakness of the deductive approaches. They are valid only if what the analyst assumes to be true and relevant actually is (Robertson 1976, p. 71). In the case of the RILE index, for example, it has been suggested that the assumption of an invariant dimension is an issue (Benoit and Laver 2006, pp. 68-69) and indeed research has shown that the meaning of left and right in general can and does change across countries and over time (Dyrberg 2009; Benoit and Laver 2006; Huber and Inglehart 1995). 3.3.3 Inductive Approaches to Positioning In order to avoid many of the controversial assumptions of deductive positioning, a number of inductive methods that aim to extract dimensions from raw data have been developed. Most of these have relied on the manifesto content analysis data. Robertson (1976) used factor analysis to explore the underlying dimensionality of content analysis data and reported seven dimensions that were worth interpreting. Also Budge, Robertson and Hearl used factor analysis to explore the political spaces of countries included in the first version of the manifesto dataset (Budge, Robertson, and Hearl 1987). More recently, Albright has used factor analysis to explore the dimensionality of the whole dataset (Albright 2010) and has noted that there is no evidence that a single dimension would be adequate to describe variation in the manifesto content analysis data. Inductive approaches have also found application in computerised content analysis. The second most known computer content analysis method, 10

Wordfish (Slapin and Proksch 2008), is in principle an semi-inductive method. It uses a set of texts and their word frequencies to estimate positions on an underlying dimension that differentiates best between the texts. The method is configured to assume one dimension, but in principle it could detect more dimensions (ibid., p. 720) and can thus be considered a semiinductive method. Inductive methods try to make as few assumptions about the dimensionality of political space as possible. The trade-off is that the results of analyses might often not be what is hoped for in terms of clarity and interpretability. It has been noted that it might be difficult to make sense of the dimensions that have been extracted from the data (König, Marbach, and Osnabrügge 2013, p. 4; Benoit and Laver 2012). This was also noted by Robertson who performed some of the first inductive analyses of party policy, but he also raised the question of why the analyst should need to interpret the content of the dimensions or parties locations on them in the first place (Robertson 1976, pp. 70-71). Furthermore, acknowledging a highly multidimensional political space complicates further analysis. 3.4 Applications of Party Policy Measures Research that applies measures of party policy can be divided into descriptive and explanatory. The first aims to characterise parties or party systems. The second to explain or predict phenomena that are related to party policy, like government formation, duration or policy output. These are the purposes of party policy measures, which should constitute the basis from which they are developed or assessed. 3.4.1 Characterising Party Systems The most common way to use the idea of party policy positions to characterise a party system as a whole is through the concept of party system polarization, which refers to ideological distance between parties (Sartori 2005; Sani and Sartori 1985). The concept of party system polarization was put in the limelight by Giovanni Sartori (Sartori 2005), for whom it was a crucial element of party systems, with high polarization being anti-thetical to democracy in a fragmented system (Sani and Sartori 1985, pp. 335-336). Dalton s recent analysis (2008) exemplifies party system polarization research. He uses voter perceptions to determine party system polarization (ibid., p. 900) and his analysis is based on the classical texts by Sartori (Sartori 2005) and Downs (Downs 1957) and thus seamlessly adopts the spatial metaphor of party competition and the left-right dimension to characterise that space. He calculates a weighted index of polarization taking into account both the ideological distance between parties as well as their size. Measures of party system polarization all tend to rely on distances on the 11

left-right scale (e.g. Alvarez and Nagler 2004; Klingemann 2006). Party system polarization has been used to explain voting behaviour and turnout (e.g. Dalton 2008; Lachat 2008) as well as other phenomena of interest in comparative politics (e.g. Wang 2012). Therefore, although it is arguably fundamental to party system description, it is also a crucial variable in the explanation of many others. 3.4.2 Explaining Party Interaction Party policy has also been a key variable in policy based theories of coalition formation (Laver 1998; Martin and Stevenson 2001; De Winter 2002) 4. Furthermore, it has been more broadly applied to the analysis of the lifecycle of governments as a variable that can be potentially related not only to coalition formation, but many other aspects of democratic governance (Müller, Bergman, and Strom 2008). Predicting coalition formation, concisely summarised by Martin and Stevenson (Martin and Stevenson 2001), has used policy as a variable since the introduction of policy-based theories of coalition formation in the 1970s. These theories rely on the spatial metaphor and on the idea that parties that are closer to each other in policy space should be more likely to form coalitions. Martin and Stevenson use the RILE index as a measure of party positions (ibid., p. 39). In addition to coalition formation, policy variables have been used in finding explanations for various different parts of the coalition government life-cycle. Müller and others use expert data and the RILE index to construct measures for the range of preferences of parties, the number of issues and several other ideological characteristics of coalition and opposition (Müller, Bergman, and Strom 2008, pp. 97-100). Most of the policy preference variables used (ibid., p. 112) are operationalised through some kind of distance between the positions of parties. These variables are used in models predicting all aspects of a coalition life-cycle. 3.5 Problems of Measuring Difference through Location Thinking of party policy is dominated by the spatial metaphor and this in turn determines not only the measurement of policy, but also the applicability of these measures. Thus, in measuring party policy it is the central task to determine the locations of parties on one or more policy dimensions that define policy space. That parties have locations in space and that there is a continuous distance between these locations is prescribed by the 4 In the context where coalition formation has become an established and respectable sub-field in comparative politics, it is worth remembering that for Downs coalition formation was in the realm of intralegislature intrigue (Downs 1957, p. 156) and thus not worth serious scientific consideration. 12

metaphor. In much of the empirical research, however, what is crucial is how different parties are from each other. Thinking in terms of locations and distances is only one possible way to interpret difference. To get to difference through position, though, brings along all the issues of the spatial metaphor and the inductive and deductive ways of interpreting policy space. To summarise the above, these include: Determining the nature of political space. Does it have the same structure and content across countries and time? Can we justifiably assume that it is a continuous space? A priori determining the number and content of dimensions. There is no guarantee that the assumptions are valid across cases, as is evident for the left-right dimension. Inductive extraction of dimensions and post hoc interpretation. Many dimensions are needed to account for the variance in the data (e.g. Robertson 1976; Albright 2010). Exactly how many are needed to adequately estimate the difference between parties is difficult if not impossible to determine and the dimensions might be impossible to make sense of. The problem of frames of reference. Making assumptions about the nature and content of political space and other things that are necessary to obtain final estimates ultimately provides results against a frame of reference constructed by the analyst, which might not only be too complex to be meaningful, but which might not be shared by the actors whose behaviour we are interested in explaining and understanding and thus undermines empirical validity. 4 Proposal for a Measure of Programmatic Overlap A measure of policy that would avoid the above-mentioned issues of the spatial metaphor, be applicable in broad cross-country time series analysis and be intuitively meaningful and understandable would thus have to satisfy the following criteria: It would have to be based on textual data as this is the only source of data that is suitable for past analyses of party policy. It would have to abandon the spatial metaphor as the foundation for thinking about policy and difference, including: 13

Assuming that the policy space is made up of spatial dimensions where parties are located as points and where political differences can be assessed as distances between points. Relying on a frame of reference with regard to the content and nature of these spatial dimensions that is constructed by the analyst and the meaning of which is indeterminate or potentially not shared by the actors in the party system. The main questions that have to be answered for an alternative measure of policy overlap are thus: How to estimate the political or ideological differences between parties directly and not relying on the spatial metaphor? How to do this without an external frame of reference, but in relation something for which we can reasonably assume that it is shared by the actors in the system? It will be argued next that a measure based on party manifesto content analysis data would fill the criteria of scope and interpretability, avoid the above-mentioned issues and provide an answer to these two questions. 4.1 A Measure of Programmatic Overlap Based on Manifesto Data As was noted above, textual data is the only kind of data that is suitable for analysing party policy over a wider scope of time and countries. Party manifestos are the best documents for this purpose. Therefore, raw manifesto content analysis data seems well suited for evaluating the political differences between parties. The most recent version of the dataset (2012b) covers 55 countries, 623 elections 905 parties and 3611 manifestos. This is the maximum extent for which it is possible to calculate the measure of programmatic overlap that is proposed here and therefore also the maximum scope for the empirical analyses that are suggested below. It should be noted that being associated with saliency theory of party competition (Robertson 1976; Budge 2001, pp. 78-85), it has been suggested that the data does not measure position and might thus not be a suitable for analysing the differences between parties (Dalton 2008, p. 904). This might not be a fair critique, since all except one category of the coding scheme (general economic goals) are explicitly or implicitly positional (Benoit and Laver 2006, pp. 65-66). Whereas positional categories directly indicate disagreement, implicitly positional categories can be interpreted to show a divergence of interests. It can be assumed there there is more similarity between parties that emphasise the same positions than there is between parties that emphasise diverging or directly oppositional positions. Thus the data is 14

suitable for estimating policy similarity/difference conceptualised as the overlap between the programmatic positions of parties. If a party manifesto represents the overall programmatic position of a party and the content analysis of manifestos as provided by the manifesto data gives the breakdown of a manifesto across different political positions, then the overlap between parties can be estimated in two ways, depending on how many parties are simultaneously under focus: For a party pair: calculating the overlap as the sum of the smallest values for each of the 56 coding categories pairs. For a set of parties: calculating the overlap as the sum of the 56 coding categories with the smallest values in the set. These provide simple and understandable measures of how different in terms of their programmatic positions parties are. Furthermore, using parties themselves as points of reference, no external points of reference assumed and determined by the analyst are needed as parties are directly compared to each other. Neither is weighting of the 56 coding categories necessary (done for example for expert data by Benoit and Laver (2006)), as it can be reasonably assumed that positions receiving more attention are also more important for the party. It should be brought out that measures similar in idea have been proposed in previous research. Sani and Sartori (Sani and Sartori 1985, p. 321) use the idea of overlap on the left-right dimension to calculate the similarity between parties. Budge and Farlie (1985, pp. 275-276) have proposed a measure of partisanship that is very close to the proposed measure of difference. They suggested using the frequencies of coding categories in the manifesto data to calculate the odds that one is looking at one rather than another manifesto. 4.2 Future Perspectives Computerised Measures of Overlap Both of the mentioned computerised measures of party policy, Wordscores (Laver, Benoit, and Garry 2003) and Wordfish (Slapin and Proksch 2008), rely on word frequencies to estimate positions. Both do this in a way, however, that makes their wider application problematic. Wordscores requires prior determination of the positions of reference texts. Wordfish uses a set of texts to estimate the positions of individual texts on a dimension or dimensions assuming (ibid., p. 719) that across the texts the political vocabulary stays the same. However, it might be possible to calculate a measure of difference by relying on word frequencies alone, which can be easily determined by automated computerised analysis. Part of the basic logic could be the same as 15

estimating difference described above. Such a method, however, would have to be modified to take into account the large number of inevitably recurring words that bear no political significance and do not meaningfully differentiate between texts. Nevertheless this would be a potentially fruitful avenue to purse for estimating this kind of difference and it would significantly broaden the range of application for such a measure of difference. 4.3 Limitations of the Proposed Measure Despite the fact that the proposed measure would avoid many of the issues that undermine the empirical validity of the available measures, it does have some limitations that have to be taken into account. There are the following: The validity of the measure depends on the assumptions that election manifestos reflect the actual political profiles of parties. The measure is most valid only for the analysis of party interaction and will not be able to tell us anything about the perceptions of voters as the latter can be assumed to be unfamiliar with party manifestos. Calculated from the manifesto data, the measure shares all the methodological issues of the data, like certain suggested issues with reliability (Mikhaylov, Laver, and Benoit 2012). Despite these issues it is the position here that given the data that is available and the measures that have thus far been used, the proposed measure is both theoretically and methodologically superior. 5 Applications I: New Perspectives for Characterising Party Systems The first of two empirical parts of the project will focus on how the proposed measures of difference can be developed into indicators for the characterisation of party systems as a whole. 5.1 Measuring Overlap Instead of Polarization The concept of party system polarization was discussed above (Sartori 2005; Dalton 2008) and thus it will not be repeated here. Suffice it to say that it can be conceptualised an indicator that expresses the weighted ideological distance between parties in a system (Dalton 2008). An equivalent measure of difference should thus be: for the system as a whole, and 16

taking into account both the programmatic differences between parties as well as their relative size. The size of parties can be operationalised as their seat shares in parliament. Determining a value for the policy part of the measure, however, is more problematic, because unlike on the left-right scale, there is no straightforward numerical value describing the location of parties. The total overlap between a set of parties is not applicable, since it considers all parties equally, thus it must be based on the pairwise measure. One way to calculate such a measure is: 1. Calculating how much overlap there is on average for each party vis-àvis all other parties. Thus we would give a measure of average overlap for each party. 2. Weigh the average overlap for each party with the parliamentary seat share of the party. 3. Add up the weighted overlaps. Thus, we would have a measure that could range from 0 (no overlap in the party system) to 1 (total overlap). It can thus be applied in party system research equivalent to measures of polarization. The proposed project would use this measure of party system programmatic overlap in the following way: 1. Calculate measures of polarisation according to the formula provided by Dalton (2008) and party system programmatic overlap as described above based on the latest version of the manifesto dataset (version 2012b). 2. Explore the relationship between the measures by correlating them. 3. Comparing the two measures as predictors of voter turnout following the model provided by Dalton (ibid., p. 915). 5 Thus, it will be possible to evaluate the measure as it performs in comparison to current measures of party system polarization. 5.2 Measuring Party System Policy Change A measure of programmatic overlap could also be adopted as a measure for policy change. Parties at one election can be compared with themselves in the previous election and the amount of overlap for each party can be aggregated into an overall measure of programmatic change in the system. This would be a needed supplement to measures of party system volatility (Pedersen 1979) and can be calculated more specifically in the following way: 5 Although it is questionable how voters might perceive the programmatic overlap measured thus. 17

1. Calculating a measure of overlap for the same party in consecutive elections. 2. Weighing the overlap with the parliamentary seat share of each party. 3. Adding up the weighted overlaps. This would give an overall weighted measure of policy change. Lower values would indicate more change and higher values more stability in programmatic positions. Since this would be a novel measure in kind, the project would propose to do the following: 1. Calculate a measure of programmatic change for all consecutive pairs of elections that are covered by the manifesto dataset. 2. Give and overview of programmatic change across countries, regions and time. 6 Applications II: Understanding Coalition Formation The second of the two empirical parts of the project will constitute an application of a measure of programmatic overlap in coalition research. It will focus on two phenomena, first explaining coalition formation in general and second explaining the formation of politically non-homogeneous coalitions. 6.1 Coalition Formation and Programmatic Overlap There have been numerous studies that have tested the effect of the relative political locations of parties on predicting coalition formation. For the sake of reference, the following will refer to Martin and Stevenson s study (Martin and Stevenson 2001), which is perhaps one of the most encompassing studies. The role of policy in their study (ibid., p. 35) is conceptualised as the political differences and gaps between coalitions partners. The more politically similar a coalition, it is hypothesised, the more likely it is to form. The variables of policy that they include are all operationalised as the absolute distance between parties using the RILE index (ibid., p. 39). Using the complete range of independent variables provided by Martin and Stevenson as a point of departure and the data that is available from the European Representative Democracy Data Archive (ERDDA), which covers more that 600 European governments in the post World War II period and includes contextual and institutional data, the project would propose to: 1. Construct a model predicting government formation that would include in addition to policy variables other significant variables identified by Martin and Stevenson. 18

2. Replicate the analysis using the RILE index based policy variables. 3. Replicate the analysis by replacing the RILE based variables by measures based on programmatic overlap. The measures could be operationalised and replaced as follows: (a) Median party in coalition. The party with the largest average overlap with other parties. (b) Minimal connected winning coalition. A coalition where the formateur has chosen as coalition partners parties that have most programmatic overlap with it. (c) Ideologically compact minimal winning coalition. A minimal winning coalition with the largest possible programmatic overlap between coalition partners. (d) Government ideological divisions. Degree of overlap between coalition partners. (e) Opposition ideological divisions. Degree of overlap between opposition parties. Using the RILE index based variables or programmatic overlap based variables in models that otherwise are exactly the same would enable to evaluate which ones are better predictors of coalition formation. 6.2 Explaining the Political Homogeneity of Government Coalitions Finally, the proposed project would focus on a particular kind of coalition government a non-connected coalition (e.g Andeweg 2011, p. 192). Conntectedness makes sense only in the context of the spatial metaphor of party competition, as it is a coalition that contains no ideological gaps, usually on the left-right dimension. This type of coalition runs counter not only to policy based coalition theories, but also the imperatives of representative democracy, which prescribe that parties should stand in government for the policies that they have advocated and thus seek cooperation with politically more similar than different parties. In the framework of the spatial metaphor, it is easy to conceptualise and operationalise a non-connected government. Using the measure for programmatic overlap it is more complicated, but completely feasible. The same measures as were described in the previous section about non-connected governments could be used. It would then be possible to conduct the following analyses: 19

1. Comparing the proposed measures of political homogeneity of coalitions to existing measures of non-connected governments, like the one included in the ERDDA. 2. Using the information available in the ERDDA about the context and process of coalition formation, test if there is systematic variation in conditions under which non-homogeneous coalitions form. From the perspective of representative democracy the important question is not if governments form (they do), but what kind of governments form under what circumstances. 7 Plan of the Project The proposed project will be completed according to the four main sections that have been presented here and in the same order. Each of the sections, that correspond to chapters of the final thesis, can be envisioned as potentially stand-alone arguments with the empirical sections building on the first two chapters. Upon completion, they could be re-written and published as peer-reviewed journal articles. The completion of the thesis does not depend on separate fieldwork and any information that would be required the for empirical analyses described in the last two sections that is not already available from existing systematized data sources like the ERDDA can be collected by the author from other sources. Thus, there are few external risks for completing the project. The time-frame for the project can be proposed as follows, assuming that it would take approximately 36 months starting from academic year 2013/2014 to complete the project: Part 1. Theoretical and empirical background of measuring party policy, including the spatial metaphor and its implications. This would be the basis for a proposal for the alternative measure. Time: 8 months Part 2. Elaboration of the new measure, including the ways it can be calculated on the basis of manifesto data. This includes calculating the measure on the basis of the data that is available from the manifesto dataset and a compilation of the corresponding dataset. Time: 8 months The first and second parts of the thesis form an integrated whole and thus they could be developed into a single foundational article that would make a comprehensive argument for the proposed measure. 20

Time: 2 months (given that parts 1 and 2 are completed) Part 3. Characterising party systems. Conducting empirical analyses related to party system polarization and proposing a measure of policy change. Given that these are separate measures and separate analyses, they could potentially be developed into two separate articles. Time: 7 months Part 4. Predicting coalition formation and explaining non-connected coalitions. Likewise, these are separate analyses and they could potentially be developed additionally into two separate articles. Time: 7 months Finalising the four parts into a completed thesis. Writing the introduction and conclusion. Time: 4 months Envisioned thus, the total approximate time to complete the project, given that it would have to be done in parallel with other activities like teaching assistantship and coursework, would be 36 months. References Albright, Jeremy J. (2010). The multidimensional nature of party competition. In: Party Politics 16.6, 699 719. Alvarez, R. Michael and Jonathan Nagler (2004). Party System Compactness: Measurement and Consequences. In: Political Analysis 12.1, pp. 46 62. Andeweg, Rudy W. (2011). From puzzles to prospects for coalition theory. In: Puzzles of Government Formation. Coalition Theory and Deviant Cases. Ed. by Rudy W. Andeweg, Lieven De Winter, and Patrick Dumont. Routledge, 191 203. Andeweg, Rudy W. and Jacques J. A. Thomassen (2005). Modes of Political Representation: Toward a New Typology. In: Legislative Studies Quarterly 30.4, 507 528. Bakker, Ryan (2009). Re-measuring left right: A comparison of SEM and Bayesian measurement models for extracting left right party placements. In: Electoral Studies 28.3, 413 421. Bakker, Ryan et al. (2012). Measuring party positions in Europe: The Chapel Hill expert survey trend file, 1999 2010. In: Party Politics. Benoit, Kenneth and Michael Laver (2006). Party Policy in Modern Democracies. Routledge. 21

Benoit, Kenneth and Michael Laver (2012). The dimensionality of political space: Epistemological and methodological considerations. In: European Union Politics 13.2, 194 218. Beyme, Klaus Von (1985). Political Parties in Western Democracies. Grower Publishing Company. Budge, Ian (2001). Theory and Measurement of Party Policy Positions. In: Mapping Policy Preferences. Estimates for Parties, Governments, Electorates 1945-1998. Ed. by Ian Budge et al., 111 126. Budge, Ian and Judith Bara (2001). Introduction. Content Analysis and Political Texts. In: Mapping Policy Preferences. Estimates for Parties, Governments, Electorates 1945-1998. Ed. by Ian Budge et al. Oxford: Oxford University Press. Budge, Ian and Dennis Farlie (1985). Party Competition - Selective Emphasis or Direct Confrontation? An Alternative View with Data. In: Western European Party Systems. Continuity and Change. Ed. by Hans Daalder and Peter Mair. London: SAGE Publications. Budge, Ian and Hans-Dieter Klingemann (2001). Finally! Comparative Over-Time Mapping of Party Policy Movement. In: Mapping Policy Preferences. Estimates for Parties, Governments, Electorates 1945-1998. Ed. by Ian Budge et al. Oxford: Oxford University Press. Budge, Ian, David Robertson, and Derek Hearl (1987). Ideology, Strategy and Party Change: Spatial Analyses of Post-War Election Programmes in 19 Democracies. Cambridge University Press. Budge, Ian et al., eds. (2001). Mapping policy preferences: estimates for parties, electors, and governments, 1945-1998. Oxford: Oxford University Press. Daalder, Hans and Peter Mair, eds. (1985). Western European party systems: continuity & change. Beverly Hills, Calif: SAGE Publications. Dalton, Russell J. (2008). The Quantity and the Quality of Party Systems Party System Polarization, Its Measurement, and Its Consequences. In: Comparative Political Studies 41.7, 899 920. Dalton, Russell J., David M. Farrell, and Ian McAllister (2011). Political Parties and Democratic Linkage: How Parties Organize Democracy. Oxford University Press, USA. De Winter, Lieven (2002). Parties and Government Formation, Portfolio Allocation, and Policy Definition. In: Political Parties in the New Europe: Political and Analytical Challenges. Ed. by Kurt Richard Luther and Ferdinand Müller-Rommel. Oxford: Oxford University Press, 171 206. Dinas, Elias and Kostas Gemenis (2010). Measuring Parties Ideological Positions With Manifesto Data A Critical Evaluation of the Competing Methods. In: Party Politics 16.4, 427 450. Downs, Anthony (1957). An Economic Theory of Democracy. New York: Harper and Row Publishers. 22