Salient Unemployment and the Economic Origins of Party-system Fragmentation: Evidence from OECD 1

Transcription

1 Salient Unemployment and the Economic Origins of Party-system Fragmentation: Evidence from OECD 1 Konstantinos Matakos 2 K.K.Matakos@warwick.ac.uk Dimitrios Xefteris 3 xefteris.dimitrios@ucy.ac.cy Job Market Paper (This version: October 31, 2011) 1 The authors would like to thank the participants at the CAGE Work in Progress Workshop at Warwick University, for useful comments and suggestions. Konstantinos Matakos would also like to thank Bhaskar Dutta and Ben Lockwood for invaluable support and supervision and Fabian Waldinger, Sharun Mukand, Philip Reny and Michail Rousakis for useful suggestions and interesting discussions. Konstantinos Matakos gratefully acknowledges nancial support from the "A.G. Leventis Foundation". All errors remain ours. 2 Corresponding author: Konstantinos Matakos, University of Warwick, Department of Economics, Social Studies S2.94, Coventry, CV4 7 AL, UK. 3 Department of Economics, University of Cyprus, Nicosia.

2 Abstract We explore the impact of unemployment on party-system fragmentation and the political institutions of industrialized countries via economic (redistributive) voting. Employing a four-party model with two dimensions of choice ( scal policy and ideology), we document a non-monotonic relationship between unemployment and fragmentation that traditional voting theory cannot account for. In equilibrium, big parties woo the group of voters (unemployed) who are relatively more willing to switch their votes in response to more redistribution(or higher public spending) nanced through labour taxation. For low levels of unemployment that allow for higher redistribution, an increase in unemployment increases the amount of "swing-voters" that are up for grabs. Hence, fragmentation initially decreases before taking the uphill. We test our ndings on OECD economies, using oil price shocks as an instrument for unemployment. The results yield robust support for our theory. We nd that variation in unemployment alone can account for two-thirds of the variation in party-system fragmentation (an 1% increase in the unemployment rate is associated with a 2% decrease in fragmentation). In fact, the economy is the single most important determinant of party-system structure in a democracy. We also provide an alternative test using data from greek regional elections to exploit the information shock on the true level of debt as a natural experiment. We nd that public spending cuts caused an increase in fragmentation. We con rm again the non-monotonic relationship between unemployment and electoral fragmentation, as public spending became scarcer. Hence, our model explains how parties preferences for scal discipline or redistribution vary with the degree of salience of unemployment. It can also explain why political systems with similar characteristics exhibit di erential degrees of tolerance for unemployment. Finally, it highlights the importance of institutional and scal constraints in guaranteeing political pluralism. Keywords: party-system fragmentation, unemployment, clientelism, public spending, redistributive voting, instrumental variables, institutional constraints. JEL classi cation: C22, C23, C26, C70, D72, H00, H69, J68

3 1 Introduction This paper explores the impact of unemployment on the fragmentation of the party-system in advanced industrialized democracies. More speci cally, we will document, both theoretically and empirically, the existence of a non-monotonic relationship between the unemployment rate and the degrees of party-system and electoral fragmentation 1. In fact, we will show that the variation in a single key economic variable, the unemployment rate, can account for two-thirds of the variation in electoral and party-system fragmentation, controlling for a set of relevant political, economic, institutional (e.g. electoral rule) and historical parameters. Furthermore, we also explore the links through which this e ect operates, o ering two alternative, but non-rivalrous, explanatory stories: one of simple economic (opportunistic) voting and another one that rests on the existence of patron-client relations, underlining the political system. Yet, irrespective of the chosen link, we document the importance of economic conditions as determinants of the party-system structure. Moreover, we highlight the role of institutional constraints in shaping the political landscape. The role of economic conditions in determining electoral fragmentation 2 and the overall structure of the party-system is of central interest for anyone trying to understand how political and partysystems operate. Increased electoral and party-system fragmentation can be viewed as a sign of increased political pluralism (more parties sharing the votes more equitably). But, at the same time, it can also constitute a sign of increased political tension, polarization and centrifugal forces operating within a society (Esteban and Ray, 2011). Since welfare and social order critically depend on political stability, the determinants of party-system fragmentation seem to matter a lot. A quick look at Fig.1, in the Appendix, shows that the level of electoral fragmentation in almost all OECD countries follows a pattern that resembles that of a political cycle. Secondly, whereas in most western societies electoral fragmentation seems to be increasing, there are some other countries, notably the south European states, where the trend of fragmentation seems to be reversed or stable. In summary, during the second half of the twentieth century, there is a wide variation in the degree of electoral fragmentation, even among countries with similar economic, socio-demographic and political characteristics. But then, why in some countries, there is only a small number of political parties (usually two) that receive the lion s share of votes, while in some others the votes are shared more equitably among many more parties? Why do we observe such a big variation in the fragmentation of the party-system, even among the industrialized OECD economies? Moreover, what are the determinants of party-system fragmentation and what is the role of economic conditions in explaining this observed variation? In particular, can changes in unemployment account for some of this variation? And through which links does this e ect operate? Is it simple opportunistic (economic) voting or can patron-client relations provide an alternative explanation? What is the role of institutional constraints? The detailed answer to those questions will be the main focus of this paper. To do so, we develop a rational choice-based model of electoral competition. We also o er some useful insights for the evaluation and design of political institutions. 1 We adopt the standard convention in the literature (e.g. Persson & Tabellini, 2007) and we measure electoral fragmentation in a way analogous to market fragmentation, using an adjusted Her ndahl-hirchman Index (HHI) where we replace market shares with parties vote shares. 2 The terms party-system fragmentation and electoral fragmentation will be used interchangeably hereinafter in this paper. 1

4 1.1 Main Hypothesis: Economic Voting and Political Unemployment The exact nature of the relationship between unemployment and electoral fragmentation seems to be somewhat puzzling. Looking again more closely at the stylized evidence, in Fig. 2 in the Appendix, we can observe a certain pattern between electoral fragmentation and unemployment. The data seem to suggest a non-monotonic relationship between unemployment and electoral fragmentation, both across and within countries. Moreover, the relationship is negative for low levels of unemployment. Initially, this might come as a surprise, since one would expect a positive relationship between unemployment and support for marginal parties which entails higher fragmentation. It would appear more likely that voters punish the larger parties when unemployment is high (retrospective voting). Nevertheless, there seems to be something more than retrospective voting going on. In order to explain what appears to be a paradox, at a rst sight, we put forward two alternative, yet complementary, stories: the main economic voting hypothesis and clientelism. Without discarding the merits of retrospective voting, we hypothesize that the relationship between unemployment and electoral fragmentation might be non-monotonic. For reasonably low 3 and intermediate levels of unemployment, an increase in the unemployment rate can lead to a decline in electoral fragmentation. This can be attributed to one or both of the following two reasons: the relatively stronger rent-seeking behavior of the unemployed voters that are attracted by generous policy promises made by big parties, and(or) the patron-client structure of the political system. But, as the unemployment rate gets out of control, the trend is reversed and higher unemployment has now the opposite e ect. Hence, it is reasonable to expect a non-monotonic relationship between unemployment and electoral fragmentation. Our main hypothesis attempts to explain this puzzling relationship by focusing on the voting behavior of the unemployed voters. In fact, we argue that the least well-o voters, the unemployed ones, are more likely to vote for the dominant parties 4, conditional on those parties credibly promising high public spending, nanced through higher taxes. The intuition behind this hypothesis is relatively simple and follows a carrot-and-stick type of argument. Unemployed voters expect higher income transfers (rents) via increased public spending, if they vote for the party that can credibly promise it and this party has good chances of winning and implementing it. Even more so, when public nances are healthy (or scal and institutional constraints are low) and allow for more generous, yet credible, public spending promises. Furthermore, voters with more humble economic means, and amongst them certainly the unemployed population, exhibit a stronger rent-seeking behavior relative to employed voters with higher incomes. Thus, ceteris paribus, they are more likely to vote for the party that credibly promises more generous public spending 5, even if this party is ideologically distant to them. This is so, because the marginal rate of substitution of income for ideology is relatively larger for unemployed voters which implies that they are willing to sacri ce more in ideology for a given increase in income. In contrast with them, employed voters would have voted for another party, given the same ideological distance. Hence, the voting behavior of unemployed voters is dominated by their 3 In the empirical section of the paper we will become more explicit as to what we consider reasonable levels of unemployment. 4 Dominant parties are those that have a strictly positive probability of winning elections. Hence, they are the only ones that can credibly promise high public spending. In presenting our model we will not assume ex-post that some parties are dominant, rather we will formally show that in equilibrium only some parties (the centrist ones) can have a positive probability of winning. Thus dominant parties will arise endogenously. 5 As we show in the next section of this paper only some parties promise high public spending in equilibrium. We call them, ex-post, dominant. 2

5 rent-seeking desire and given su cient transfer promises ( carrot ), unemployment acts like the stick. The idea of unemployment being used as a coercion or discipline device is not new in economics. Shapiro and Stiglitz apply this idea in the labour market. Yet, our application on political theory and voting behavior is novel. If Tullock was right in asserting that voters and buyers are essentially the same people, then it is normal for us to go one step further and treat parties like rms that try to increase their vote (market) shares and dominate the electoral (market) competition. Therefore, paraphrasing K. Marx 6, we hypothesize that big parties require a reserve army of (unemployed) voters in order to dominate in the political competition. Nevertheless, this argument has its limits. When unemployment gets out of control, inequality increases and the tax base shrinks dramatically, thus making income transfers virtually impossible. As a result, big parties can no longer exploit the rent-seeking behavior of the unemployed who now vote against them. We have assigned the term Political Unemployment to characterize this situation where dominant parties exploit salient unemployment for own electoral bene t. Clearly, it portrays the dilemma which unemployed voters are caught in. Instead of punishing the dominant parties for their policies that have led them into being unemployed, they are more likely to vote for them since they are dependent on their transfers. The reason is the rent-seeking behavior dominates their voting decision. Hence, as the unemployment rate increases, so does the vote share of dominant parties. Another potential, non-rival explanation could be the existence of clientelistic relations in the political system. That is, big political parties (patrons) can buy the support of the unemployed voters by promising them various forms of nancial assistance, such as public sector jobs and all sorts of favorable, non-meritocratic arrangements. In our study, Greece and its clientelistic party-system will be used as a test-case for the latter hypothesis. In the rst place, our claim might sound a little counter-intuitive, given the traditional approach of retrospective voting. Yet, it is in line with rational voting theory. Moreover, if it is empirically con rmed, it paves the ground to complete and revise the theory of political business cycles (Nordhaus 1975; Alesina et al. 1987, 1992) in two directions: Firstly, it will complete the two-way relationship between economic and political outcomes. So far, in the existing literature (Barro 1996; Persson et al. 2007) the impact of political institutions, such as the electoral rule, and political outcomes on real economic variables has been extensively studied and very well documented both theoretically and empirically. Nevertheless, the reverse direction of causality is understudied. Hence, our study aspires to bridge this gap by providing empirical and theoretical evidence on the e ect of unemployment on party-system structure. Secondly, our hypothesis can provide a political raison d être for unemployment above the level that is known as the natural rate. From the above discussion, it has become evident that in the pre-election years dominant parties would have an opportunistic incentive to pursue policies that favour scal discipline and prudence (e.g. scal rules, balanced-budget amendments and other institutional constraints) even at the cost of slightly higher unemployment since then, they can exploit this reserve army of potential voters for own bene t. Thus, our framework allows for endogenizing political parties preferences over unemployment in a way that the partisan political business cycles theory is unable to capture. Finally, in parallel with our main hypotheses, we put forward an additional claim that stems out from observing the data on Greek Regional Elections (Fig. 3). That is, we observe that lower 6 Karl Marx s initial quote was "...Big industry constantly requires a reserve army of unemployed workers..." 3

6 expected public spending is associated with higher fragmentation. Here we exploit the revisions 7 on the level of greek public nances, for the period, as an information shock 8 that generated expectations for a dramatic cut in public spending. The intuition is clear: the realization of scal derailment and the resulting austerity measures limited, in turn, the ability of the big parties to credibly promise higher public spending in the future. These expected cuts severely weakened the strength of clientelistic links, thus undermining the bipartisan foundations of the greek party-system, since there is not much left for the parties to promise anymore 9. As a result, expected public spending cuts mitigated the rent-seeking behavior of unemployed voters who are now less likely to vote for the dominant parties, ceteris paribus. This in turn led to a record increase in electoral fragmentation, as evidence from greek local elections suggests. Yet, contrary to our Economic Voting hypothesis, this latter claim does not come as a surprise since historically economic adversity and low public spending are associated with an increase in both electoral fragmentation and polarization (e.g. Persson & Tabellini, 2007). For this reason, the primary focus of our study is the political unemployment paradox. 1.2 Literature Review Political scientists have always been interested in examining this phenomenon and have adopted various approaches to explain it. Duverger (1954), rst noted that the observed variation in electoral fragmentation among countries may be attributed to the particular political institutions that characterize the political environment of each state, the most prominent of those institutions being the electoral rule. Duverger s rst law asserts that "the simple majority, single ballot system favours the two-party system". That is, if a majoritarian electoral rule is applied, then we should expect a low level of electoral fragmentation, where the two larger parties should be receiving the larger share of votes. By contrast, his second hypothesis 10 suggests that "both the simple-majority system with second ballot and proportional representation favour multi-partism". Hence, when a proportional electoral rule is in place, fragmentation should be relatively higher. Individual voters are driven to vote for the larger parties in the rst case, as a vote for a small party in a majoritarian system is deemed as a lost vote. In a sense, this accounts to citizens voting strategically, since voting for a minor party that has minimal chances of wining may alter the election result in their disfavor. Nevertheless, this approach does not account for the complete story. Looking again at the data, we observe the following paradox: countries with extremely stable electoral rules and solid party systems throughout their recent political history, like the United States, the United Kingdom, Germany and Sweden 11 exhibit a rather unexpected volatility in electoral fragmentation overtime. Interestingly enough and contrary to Duvergerian predictions, countries like Greece, where changes in electoral rules have been frequent over the last years and the party system is more fragile, exhibit 7 As it can be seen in Table 14, Eurostat revised its estimates on the level of greek public debt upwards, twice within a year (once before the 2009 Legislative Elections and once before the 2010 Local Elections), and by more than 27% over the period extending from of which 11.7% referred to the 2006 de cit. 8 We refer to this second revision as a pure information shock since it was due to the greek government s book- ddling activity and misreporting of the de cits during the whole period from Since this was revealed on a later date (October 2010), and prior to the 2010 Local Elections, we consider it as a perfect information shock that reached the whole country but a ected the regions di erentially. Since the de cits were already there (but remained uncovered) they could not have been outcomes of the 2009 elections. Yet their revelation a ected the vote in the 2010 elections via altering voter s expectations on future public spending di erentially. 9 In this particular case due to an exogenously imposed institutional constraint (the IMF-EU bail-out agreement). 10 The division of Duverger s two statements into one law and one hypothesis is due to Riker. 11 The rst two having a majoritarian electoral rule ( rst-past-the-post), with Germany having a mixed rule and Sweeden having a PR rule with a list. 4

7 signi cantly lower volatility in electoral fragmentation. To address this issue, two strands were developed in the literature. Firstly, recent political science literature, initiated by Colomer et al. (2005), points out that the electoral rule itself might also be endogenously determined through some political processes. These in turn, might depend on electoral fragmentation. Secondly, economic outcomes might also be linked with electoral outcomes and hence electoral fragmentation. To the support of this hypothesis there is a large and growing body of literature in economics. From a purely economic perspective, several studies have been conducted that relate electoral outcomes with macroeconomic variables. Alesina (1987), Persson and Tabellini (2003), Persson et al. (2007), Acemoglu et al. (2000; 2005) and Barro (1996), have extensively analyzed the impact of institutions and other political indicators on the economic performance of a country. Speci cally, Barro studies how political development and institutions a ect economic performance and growth, whereas Alesina develops a rational expectations political cycles model, where partisan electoral competition a ects unemployment and in ation. More recently, Persson et. al (2007) examine the e ect that the electoral rules have on government spending in parliamentary democracies. They conclude and provide empirical evidence that electoral rules a ect public spending indirectly through the fragmentation of the party system. Thus, they support that "proportional elections induce a more fragmented party system". As a result, they are associated with a larger incidence of coalition governments, which in turn induce higher public spending than single party governments. In fact, their results are in line with Duvergerian approach, since they suggest a causal link between electoral rules, the resulting party system fragmentation and economic outcomes. Nevertheless, this approach explores the link between fragmentation and economic outcomes in one only direction. Namely, how the fragmentation of the party system a ects public spending. Yet, although the impact of political institutions and party-system structure on economic and political outcomes has been extensively studied (Alesina and Roubini, Barro, Persson and Tabellini, Acemoglu et al.) the reverse relationship has not been documented systematically. In speci c, the impact of unemployment on party-system fragmentation, as measured by electoral fragmentation, is a fundamental question that has not yet been thoroughly addressed. As such, it cannot explain how economic outcomes might a ect fragmentation. Therefore, we are interested in the reverse direction of the relationship between economic conditions and fragmentation. The in uence of economic variables on electoral outcomes has been studied by a large variety of scholars. Both at an individual (Fiorina 1981; Kinder and Kiewiet 1979, 1981; Kiewiet 1983) and at a countrylevel (Kramer 1971; Arcelus and Meltzer 1975; Bloom and Price 1975; Tufte 1978; Kinder Adams, Gronke 1989; Campbell 2000; Holbrook 2008), economic conditions are found to have a severe impact on determining the winner of an electoral competition. Nevertheless, the main focus of this literature is not that broad in the following sense: it might be informative on a partisan level, but it fails to capture whether economic conditions have an impact on the shape and structure of the party system itself. Moreover, it is silent with respect to the centrifugal or centripetal forces that dominate the political system and the role of institutional constraints. Our paper di ers from the existing literature in many respects. Firstly, it attempts to shed some light on the reverse relationship between economic and electoral outcomes. To this aim, we also o er two alternative explanatory links. Secondly, the scope of our study is more broad. We are interested in studying the impact of economic conditions on the structure of the party system, not just on the chances of re-elections of the incumbent party. For this reason, we focus on a broader de nition of the electoral outcome: the fragmentation of the party-system. This is so, because 5

8 electoral fragmentation can better capture the changes that occur in the party system, due to changes in the economic conditions. Finally, our study extends the results to a set of countries using cross-country data. Hence, we set out to explore the reverse causal relationship between economic outcomes and political fragmentation. And given that the key economic indicator is the unemployment rate, we will attempt to account for the variation in electoral fragmentation caused by changes in unemployment. 2 Theoretical Model In this section, we will rst solve a general theoretical model and then test its implications by the means of an empirical study. We consider that the electoral competition takes place among four parties, 12 in a two dimensional space. We will name the two dimensions ideology and public spending 13 respectively. The two dimensions di er from each other only in one respect. In the ideological dimension the positions of the four parties will be xed while in the second dimension (public spending) a strategic choice will be made by each one of them. 14 Moreover, the preferences of the voters in both dimensions will be heterogeneous (unemployed voters prefer high compared to low public spending, whereas the opposite is true for the employed ones). 15 Finally, both dimensions will be composed out of a continuum of policies. In the public spending dimension the promised level of spending will vary proportionately with the level of the tax rate, which is chosen from a continuum Political Parties After describing the nature of electoral competition, we can formally de ne the party-ideology space is as follows: 1 P = fl; L; R; rg [ ; 1 ] such that l < L < R < r 2 2 where l is a party with an extreme-left ideology, L is a party with a moderate-left ideology, R is a party with a moderate-right ideology and r is a party with an extreme-right ideology. We shall assume that a party s index p 2 P coincides with its ideological position. The ideology space is 1 the normalized interval [ ; 1 ]. In order to give more structure to our model we will make the 2 2 following assumption with respect to the parties position on the ideology space. Condition 1 (Symmetry) Parties l and r are positioned in the extremes of the ideology space, that is at l = 1 and r = 1, respectively. Parties L and R are symmetrically positioned at distance 2 2 around zero: That is, at L = and R = respectively. 12 In another version of the paper, we allow for endogenous party-entry, where four parties arise as an equilibrium outcome. 13 Persson and Tabellini (2000) 14 Krasa and Polborn (2010) consider a similar model of electoral competition where candidates...are exogenously committed to particular positions on some issues [ideology] while they choose positions for the remaining issues [public spending]. Moreover, Dziubinski and Roy (2010) consider a model with exactly two dimensions where parties are commited in one dimension but have the freedom to (credibly) choose any position in the other. 15 In an earlier version, we presented a variation of the model where voters have homogenous preferences on the public spending dimension (everyone prefers high to low public spending). 16 In an earlier version, following Krasa and Polborn (2010) we presented a model where the public spending dimension was binary. 6

9 Each party s ideological position is public information and is known to every agent in the society. Furthermore, each party will commit in the pre-election stage to a level of public spending. Since the budget must be balanced, the total amount of public spending should always equal the total revenues raised through taxation. As a result, any choice of public spending is uniquely associated with a tax rate proposal. That is, each party p can promise a tax rate t p to the citizens such that t p 2 [0; ] such that 2 (0; 1). Clearly, a proposal of tax rate is associated with highest possible level of public spending and vice versa. Further assume that each party s utility function has two components. Parties derive utility both from the ideology that is implemented, as well as from the vote share that they receive. The level of public spending does not directly a ect their welfare. 17 Hence, under this formulation parties are simultaneously ideology-driven and also rational vote-share maximizers. Formally, their utility function takes the following form: V p (!; v p ) = j! pj + b(v p ); p 2 P where v p is the vote share of party p 2 P, and! is the ideology that gets implemented by the winner, once the electoral result has been realized. Since the party that receives the larger share of votes wins the election and forms the government,! also depends on the resulting vote share allocation 18. That is, whenever v L > v p for every p 2 P n flg then, party L wins in the electoral competition and, given that its ideological position is xed, it implements! L =. Therefore, given that! is the ideology of the winner, it clearly depends on the vote share allocation. Hence, we 1 can formally de ne it as follows:!(v) 2 [ ; 1], where v is the vector of vote shares (v 2 2 l; v L ; v R ; v r ). The term b(v p ), is an increasing function with respect to v p and measures the bene t that each party derives by securing some votes in the parliamentary elections. Hereinafter, for computational ease, we assume that b(:) is linear. This formulation of the utility function allows us to capture the negative externality that is imposed to each party if a party other than itself wins in the electoral competition. It also captures the bene t that each party receives from being a recognized party that is represented in the parliament. To conclude the discussion in this section, we note that each party s ideology p, together with its tax rate proposal t p (associated with a corresponding level of public spending) comprises party p s political platform, upon which citizens vote. That is, political competition takes place in twodimensions: an ideological one (non-monetary component) and a public spending one (monetary component). Nevertheless, given that for every party ideology is xed, in practice, the game involves strategic decisions only in the tax rate (public spending) dimension. 2.2 The Voters We consider two groups, each group consisting of a continuum of voters, whose ideological preferences (ideal points) are single-peaked and distributed according to a uniform distribution on the 1 ideology interval [ ; 1]: That is, x U[ 1 ; 1 ] in both cases. The two continua are identical in all respects apart from two: rstly, they can be of unequal size and secondly, citizens have di erent initial incomes. We assume that the rst group of voters represents a fraction q 2 (0; 1) of the 17 But it does so indirectly by altering their vote shares. 18 If no party gets the absolute majority, then, a minimal winning coalition is formed and the implemented ideology! is an average of the two parties ideologies weighted by their vote shares. For example,! = v LL+v l l v L +v l ; where L = and l =

10 society. These voters will be considered to be the unemployed ones, whose income m 2 (0; M) is a fraction of the market wage M. The second group represents the remaining (1 q) of the society. These voters will be considered to be the employed ones receiving income M. Since m is a fraction of M, we can always normalize M = 1 such that m 2 (0; 1). As a result of di erent initial incomes, agents have di erential preferences with respect to taxation and public spending. Following a redistribution mechanism analogous to that of Meltzer and Richard (1981), we can compute the total revenue raised, and transferred to the citizens in the form of public spending, when a party proposes tax rate t p 6= 0. Otherwise, if t p = 0, public spending is also 0. Then, since budget balance is satis ed with equality, the total revenue raised and transferred to each voter is given by the following expression: T (t p ) = t p [qm+(1 q)m] ; t p 2 [0; ] such that 2 (0; 1) q+(1 q) Given initial income y 2 fm; Mg, with m 2 (0; M), the utility of an agent with ideal policy x and income y (we will name this voter fx; yg) is given by: U(!; x; y) = j! xj + f(y + T (t! )), where T (t! ) is the transfer proposed by the winner!. The rst component of this expression, namely j! xj, is the utility that voter x receives from the ideology of the winning party. The second component, f(y + T (t! )), is the utility that voter x is receiving from her initial income y 2 fm; Mg, plus the promised level of public spending T (t! ), given the proposed tax rate t! of the winner. We assume that f(:) is any continuous, strictly concave and twice continuously di erentiable function with f 0 (:) > 0 and f 00 (:) < 0. Then, the resulting utility of income for an unemployed voter, given transfer T (t! ), is given by the expression: f (m(1 t! ) + qmt! + (1 q)mt! ) = f (m + (1 q)(m m)t! ) > f(m) whereas, for an employed one is given by: f(m(1 t! ) + qmt! + (1 q)mt! ) = f(m q(m m)t! ) < f(m) Clearly, since the RHS of both the above inequalities is the resulting income when the tax-rate and hence, the transfer are zero, we can deduce that all unemployed voters prefer the highest possible tax rate since redistribution takes place in their favour. However, the employed ones have no preference for redistribution and strictly prefer the zero tax-rate 19. Hence, any party proposing a positive tax rate will be winning the votes of the unemployed at the cost of losing votes of the most well-o members of the society. But, given the proportional taxation scheme, parties will always have an incentive to propose a strictly positive tax rate, even for extremely low 19 This formulation of preferences, within each group, is exactly equivalent to Groseclose s (2007) one-and-a-half dimensional preferences where...alternatives [in our case parties]...are described by two characteristics: their position in a left-right dimension, and their position in a good-bad [high-low tax rate] dimension, over which voters [of the same group] have identical preferences. 8

11 levels of unemployment, thus, capturing the vote of the unemployed. The intuition is the following: If q is small the tax base is large and whatever T () is collected, regardless of how low or high might be, is distributed among fewer people. Since the marginal utility of income is higher for the unemployed, but the marginal utility of ideology is identical, it is relatively easier for the parties to capture the vote of the unemployed given some ideological distance. Hence, proposing the highest possible tax rate will always bring more votes to all parties. But then, one might ask the following question: why don t we observe all parties making identical tax proposals? The answer to this is that, as our model predicts, some parties will choose to behave strategically and try to alter the electoral outcome by undercutting their opponents. Hence, our model s predictions are in line with reality. We will come back to this point in the next section. But rst, we want to formally present and discuss our main results. 2.3 The Voting Game We consider a voting game with three stages. Ballots, in this model, are secret. All information is publicly available and known ex ante to all the agents in the society. That is, the model does not exhibit any uncertainty or informational asymmetry. The equilibrium solution concept we employ is Nash equilibrium. The three stages of the game are as follows: Stage 1: Parties announce simultaneously their complete political platforms fp; t p g, which become public information. Given that the ideology dimension of each party p is xed, we can save in notation by omitting p from fp; t p g. Since t p is the only strategic choice made by parties, we can rewrite their maximization problem as follows: max V p (t p ; t p ) = j!(v(t p ; t p )) pj + v p (t p ; t p ) t p The winning party is denoted as!. Clearly, the winner depends on the vote share that each party received during elections, which in turn depends on the tax-rate proposal t p, proposed by each party. Formally we have: v(t) 2 [0; 1] and!(v) 2 [ 1 2 ; 1 2 ]. Hence, we denote the winner of the electoral game as!(v(t)); where v is the vector (v p ) p2p and t is the vector (t p ) p2p. Stage 2: Each individual votes sincerely for her most preferred platform, given parties platform promises. Formally, sincere voting in this setup means that each voter fx; yg faces the following maximization problem: max p2p U(p; x; y) = jp xj + f(y + T (tp )): Stage 3: Given voters decisions on stage 2, each party receives its vote share v p 2 [0; 1] such that P p2pv p = 1; and the voting outcome is realized. The party that collects the majority of the votes wins the electoral competition and is called upon to form the government and implement 9

12 its political platform 20. In case of ties, the winning parties do so with equal probability. We assume commitment. That is, the winner fully implements its ideology and realizes its tax (public spending) promises. As stated earlier, the electoral outcome depends on the distribution of ideological preferences, 1 which is uniform in [ ; 1 ], on the positions of the four parties in the unidimensional ideology 2 2 space and on the tax rate (public spending) proposal of each party. Hence, parties ultimately choose t p, and all that citizens do, is voting sincerely for the party p 2 P, whose proposed tax rate t p in conjunction with its ideology maximize their utility. As a result, by construction of the model, there is no strategical dimension in the voting decision of any individual voter. Finally, to conclude the discussion, we will now de ne a key variable of interest. During the elections each party receives a vote share, denoted by v p 2 [0; 1]. Following Rae (1968) and Laakso and Taagepera (1979) we can de ne electoral fragmentation as an inverse Her ndahl-hirchman Index (HHI ): P F (v) = 1 (v p ) 2 p2p 2.4 Results In this section, we present our main results. In trying to establish our equilibrium characterization result, we note that we are not restricting attention to any particular class of Nash equilibria. In fact, we will show that for relatively mild assumptions, the game has a unique equilibrium which also happens to be symmetric in nature. Since the main purpose of this essay is to study how economic parameters a ect electoral fragmentation, a symmetric equilibrium can provide the necessary framework to perform a comparative statics analysis. Therefore, it is reassuring that uniqueness of equilibrium solves for us the equilibrium selection problem. To prove our main result, we construct our argument in two steps. First, we show how parties vote shares and chances of electoral success vary with the chosen tax-rates for all possible values of the parameters. Then, we show uniqueness by highlighting the strategic behavior of the extremist parties. Before presenting the results, we de ne a symmetric equilibrium. De nition 1 An equilibrium is symmetric if and only if parties play mirror strategies. That is, both conditions have to be satis ed: (i) t l = t r and, (ii) t L = t R : We also denote a useful function that measures the maximum gain in vote share for a party, resulting from a change in the tax-rate proposal from 0 (minimum) to (maximum). Formally, for every q; 2 (0; 1) and every m and M, such that m 2 (0; M) we let: z(q; m; ) q [f(m + (1 q)(m m)) f(m)] (1 q) [f (M) f(m q(m m))]. Then, the following Lemma will help us to shed some light on the equilibrium behavior of the parties vis-à-vis their choice of strategies (tax rates). 20 In the case that no party gets the absolute majority, a minimum winning coalition is being formed among the parties with higher ideological proximity. 10

13 Lemma 1 Let (1 q)(m m) (q; m; ) and q(m m) (q; m; ). Then, the following are true: (i) z(q; m; ) q [f(m + ) f(m)] (1 q) [f (M) f(m )] is a positive, continuous in [0; 1] and di erentiable in (0; 1) function, for every q; m; and 2 (0; 1); z(q;:) < 0 for 2 q and 8m; 2 (0; 1) and > 0 for all and 8q; m. In gures 7.a and 7.b we o er a graphical exposition of Lemma 1, for f(y) = p y. Clearly, z(:) is strictly concave with respect to q, strictly increasing with respect to, strictly decreasing with respect to m and only takes positive values. It is also bounded above (by 1 ) and below (by 0). The 4 rst component is the gain in votes from the unemployed voters for a party proposing tax rate. The second part captures the loss in votes from the employed voters 22. As a result, since z(:) is always positive, it can be interpreted as the net gain in votes for a party that proposes the highest possible tax rate. Although in the previous section, we have provided some intuition as to why a party can gain more votes if it proposes a high tax rate, even for low levels of unemployment, we postpone a more detailed discussion until the next section, after presenting our main results. First, we state our characterization result. Proposition 1 For every q 2 (0; 1), every m 2 (0; M) and every 2 (0; 1), 9 ^ 2 (0; 1 ) such that 8 2 > ^ the following vector t = (t l = t r = 0; t L = t R = ) constitutes the unique Nash equilibrium of the electoral game. Moreover, the induced vote share allocation vector v (t ) = (vl ; v R ; v l ; v r) takes the following form: vl = v R = z(q; m; ); and v l = vr = 1 1z(q; m; ) Corollary 1 The vector v (t ) of Proposition 1 induces the following outcome of the electoral game:!(v (t ) = 1 2 L R: Our rst result simply says that we can always guarantee that for a large range of values for, there exists a unique and symmetric Nash equilibrium of the electoral game such that the two moderate parties propose the highest possible tax rate, the two extremists propose zero tax, and the vote share allocation is symmetric with the moderate parties winning more votes than the extremist ones. Clearly, such an allocation of votes induces, in turn, a symmetric equilibrium on the electoral game as a whole, where the median policy is implemented. Of course, given the symmetric nature of the model, this should come as no surprise. Yet, it is interesting to explore what happens for all other values of since it might be the case that we do not have an equilibrium in pure strategies (see Figs. 15.a and 15.b). The following result generalizes our equilibrium characterization result in mixed strategies. Theorem 1 For every q 2 (0; 1), every m 2 (0; M), every 2 (0; 1), and every 2 (0; 1), 9 2 < such that the following vector (t) = ( l (t l) = r(t r ); L (t L) = R (t R)) constitutes a symmetric equilibrium of the electoral game in mixed strategies, where l and r are non-degenerate with support on [0; ] and E[ l (t l)] = E[ r(t r )] = <, whereas L, R are the degenerate strategies with L () = R () = 1. Moreover, 9 ^ 2 (0; 1 ) such that 8 > ^ the equilibrium of the game is 2 unique and = All Proofs in the Appendix. 22 Clearly measures the net income transfer to an unemployed voter, due to taxation, whereas measures the net income transfer to an employed one. 11

14 This result guarantees the existence of a symmetric equilibrium for almost all values of. Moreover, for some values of the equilibrium is unique, in which case Proposition 1 becomes a special case of the Theorem where all parties choose degenerate strategies with L (t L = ) = R (t R = ) = 1 and l (t l = 0) = r(t r = 0) = 1. Furthermore, even if the equilibrium is not unique, in any (t ) the tax-rate proposal of the two extremist parties is (in expectation) strictly less than that of the two moderate ones. The only caveat is that for each value of < ^ we would have a unique -related mixed strategy equilibrium (t ) with the above characteristics. Nevertheless, since for any such we still have that <, our comparative statics analysis would have yielded qualitatively identical results. The reason is that our obtained results depend on extremist parties proposing a lower level of tax-rate (and public spending respectively) than that proposed by the moderate ones. Since this is always true in any mixed strategy equilibrium we need not worry further. Yet, solely for expositional simplicity, we have chosen to focus on the case where > ^. We defer further discussion on mixed strategy equilibria in Section Discussion of the Results One interesting feature of our equilibrium is that moderate parties will always go after the vote of the unemployed and propose highest possible tax-rate (public spending) at the cost of losing the votes of the better-o individuals. This is so, even when unemployment is extremely low. The reason for this is that the tax base is large and for any positive tax-rate the collected amount is distributed among fewer unemployed voters as public spending, thus, maximizing its e ectiveness. Given that taxation is proportional, unemployed voters, especially if they are less in number, gain more in relative terms than what the employed ones lose. Since the unemployed have larger marginal utility of income relative to the employed ones, but identical with respect to ideology, any party proposing a higher tax rate (and public spending) su ers a small loss in votes from those better-o but gains much more from the unemployed ones 23. Hence, moderate parties are pursuing a redistribution policy for purely opportunistic, vote-maximizing reasons rather than ideology. Figure 14 compares the utility of a moderate party proposing the highest tax-rate with its utility when it proposes tax-rate equal to zero. Clearly, for every q 2 (0; 1) the moderate party is better-o by proposing the high tax-rate. Another interesting feature of the equilibrium structure is the fact that the two extremist parties are purely ideological and never win elections. That is extremist parties focus solely on the ideological dimension of their platform and propose low public spending, even if this costs them losing the elections. But note that, this is not imposed exogenously to the model. Nor does it happen because we assume ex ante that those parties have no chance to win in the elections. Rather, it is an optimal strategic choice made by those parties. Someone might think that this behavior is not intuitive, since those parties could promise high public spending, especially if they are less likely to get elected and implement their promises. In a sense, their promises are empty and hence cost-less, since they are o the equilibrium path. The reason is that if the two extremist parties propose high public spending they will be shooting themselves in the foot by reducing the vote shares of their sister-parties, thus causing another less desired party to win the election. In that respect, what appears to be a costless promise changes the equilibrium outcome to their disfavor and causes a less desired party to get elected with probability one. Hence, this deviation imposes a negative externality to them. Figure 15.a 23 To see this mathematically, check that the function z(q; m; ) is strictly positive for every q; m;. 12

15 represents this point graphically. So, our model extends the idea that there is no free lunch into the eld of political competition. Nevertheless, this implicit collusive behavior cannot o set electoral competition completely. Because, at the same time, they also reject the option of withdrawing from the competition. Participation in elections secures them a positive vote share. This allows them to enjoy a status of a recognizable party, with parliamentary representation and state nancing, able to in uence the agenda-setting process in the parliament in a way that they could not do, were they merged with the moderate parties. Hence, our model delivers some degree of collusion among parties with ideological proximity even under the most competitive conditions found in a multi-party system. That is why, we think that simple as it may appear, our model can fully capture this key trade-o that is a trademark characteristic of multi-party electoral competition Main Comparative Statics Results After this brief analysis of our equilibrium, we proceed by presenting our main comparative statics results. That is, we will examine how electoral fragmentation varies with changes in unemployment and public spending. Hereinafter, for computational P simplicity, we let > ^. First, let us compute the electoral fractionalization index F (v) = 1 (v p ) 2, as a function of the vote share allocation that corresponds to the symmetric equilibrium of Proposition 1. Recall that: p2p vl (t ) = vr (t ) = [z(q; m; )] and v l (t ) = v r(t ) = [z(q; m; )] : 2 Hence, we can compute: F (v) = 1 2[v L (t ) 2 + v l (t ) 2 ]: We can, then, rewrite F as a function of q; m and : n F (q; m; ) = z(q; m; ) o 1 z(q; m; ) 2 = 2 = z2 (q; m; ) + z(q; m; ) + ( )2 + ( 1 4 )2. 2 For simplicity let ( )2 + ( )2 = C, a constant, so that F becomes: F (q; m; ) = 1 [z 2 (q; m; ) + 2z(q; m; ) + 2C]. 24 There is a clear analogy between our model of imperfect multi-party competition and oligopolistic competition with leader- rms. 13

16 Proposition 2 Assume that conditions of Proposition 1 hold. Then 9 ~q such that F (~q; :) = 0 and the following statements are true: (i) for q 2 (0; ~q) the electoral fragmentation index F (q; :) decreases as the unemployment rate is increasing (that is < 0), ceteris paribus; and (ii) q 2 (~q; 1) the fragmentation index F (q; :) increases as unemployment increases > 0). Under this equilibrium structure the indi erent voter among parties L and R is the median, due to the fact that the two moderate parties make the same public spending (tax rate) proposals. Hence, the distribution of votes among the two moderate parties remains constant. As a result, the e ect of unemployment on the fragmentation index comes through the shift of votes from extremist parties, l and r, to the centrist ones, L and R; respectively. In that sense, we are measuring the change in the electoral fragmentation net of the incumbency e ect. Therefore, the changes in the index occurring in our model capture the net changes in the party-system structure and the movement of voters from moderate to extremist parties. That is, we disentangle the incumbency e ect (vote di usion between the two centrist parties) from the systemic e ect (vote di usion from the centre to the periphery). The comparative statics results of Proposition 2 con rm the Economic Voting hypothesis for low and intermediate levels of unemployment. They can be interpreted as saying that due to the concavity of the utility function with respect to income and public spending, unemployed voters with lower initial income are more likely, ceteris paribus, to vote for a party that o ers high public spending due to stronger rent-seeking behavior. Given that in equilibrium, only the two centrist parties promise high public spending, it follows that an increase in the unemployment rate will result in those parties increasing their vote shares. Thus, electoral fragmentation decreases. Nevertheless, as we can see in Fig. 5.a and 5.b, the relationship between unemployment and electoral fragmentation is non-monotonic. It may be the case that for low and intermediate values of unemployment the economic voting mechanism is functioning. Yet, once unemployment goes out of control, the tax base shrinks dramatically and the ability of the parties to raise revenues and promise high public spending is severely undermined. In such a case further increases in the unemployment rate result in less revenues and hence, less public spending and higher fragmentation. This nding is in line with stylized evidence showing high party-system fragmentation both in extremely developed countries with low unemployment (e.g. Norway, Finland, Sweden) and also in those with extremely high levels of unemployment (e.g. Eastern Europe). Moreover, this nding is later con rmed by our empirical estimation. Furthermore, our model can explain, via rational and opportunistic motives, why in di erent societies political parties have di erent tolerance levels for unemployment. Observe that, as shown in the proof of Proposition 2, F (q; :) is non-monotonic with respect to q. 25 For every q < ~q, we have shown that F (q; :) is decreasing with respect to q (q;:) < 0), whereas the opposite is true for every q > ~q (i.e. > 0). That is, there exist ~q(m; ) such that F (~q; :) = 0 for all m; 2 (0; 1). But, in turn, ~q(m; ) varies with m and. The following Corollary summarizes this relationship. Corollary 2 Assume that conditions of Proposition 1 hold. That is, 9 ~q(m; ) such that F (~q; :) = 0 for all m; 2 (0; 1). Then, the following statements are true: ~q(m;) < 0 and, (ii) The above Corollary just states that the critical point ~q(:), after which fragmentation F (:) becomes an increasing function of unemployment, varies with m and. It is increasing in and 25 For a graphical exposition see Figures 5.a and 5.b. 14

17 decreasing in m, as we can see in Fig. 6.a. Hence, in societies where either the tax rates (and public transfers) are low, ceteris paribus, political parties have a lower tolerance for unemployment (e.g. USA) simply because the critical-level of unemployment, above which higher unemployment rate stops to serve their opportunistic goals, is lower. Equivalently, in societies where unemployment bene ts are high (e.g. Sweden, Norway) the tolerance for unemployment is again lower, ceteris paribus. Finally, our last result summarizes the relationship between electoral fragmentation and public spending. Here, we need to clarify that public spending is endogenously determined through the choice of the tax-rate. Yet, in equilibrium, parties are making a choice whether to raise a at-rate tax or not. The comparative statics exercise, therefore, refers to exogenous variations on the maximum rate that is allowed. Proposition 3 Assume that conditions of Proposition 1 hold. Then an increase in the tax (;:) (public spending) causes a decrease in electoral fragmentation F (q; m; ). That is, < 0, all 2 (0; 1). This proposition says that as institutional or political constraints are lifted and higher tax rates are possible, higher public spending increases the leverage that moderate parties (which are the only ones that propose taxes and win in equilibrium) have on the unemployed voters and hence, it also increases their vote shares. As such, electoral fragmentation is strictly decreasing in. This time, the relationship is monotonic. Figure 6.b in the Appendix summarizes the relationships of Propositions 3 and 4 altogether. 2.5 Some Remarks After presenting our main comparative statics result, some technical remarks with respect to the equilibrium structure are now in order. First, notice that regardless of our choice of, our comparative statics results go through as presented in Section But apart from expositional simplicity, there is another more intuitive reason, related with the desired nature of political competition, that led us to focus our analysis on those particular values of. If converges to 1, this means that the two moderate parties are not so close to each other anymore. Rather 2 they are converging to the extremes in terms of ideology. But then, the two centrist parties are in reality extremists, and their ideological proximity with the extremist parties drives the latter to withdraw from the political competition. In such a case, our four-party model becomes a standard two-party model where the study of fragmentation becomes trivial. On the other hand, if converges to 0, the two moderate parties converge to the centre (and to each other) making it very hard to distinguish any ideological di erences among them. In this case, the two extremists will be less willing to sacri ce even a tiny vote share, by promising lower tax-rate (public spending), in order for their neighboring moderate party to increase its chances of winning, simply because they are now indi erent between the two moderate parties that have converged to the median. Hence, they prefer to promise high spending as well leading to an equilibrium in which the public spending dimension is practically cancelled out since all parties choose identical strategies. Therefore, in order to study how electoral fragmentation evolves when macroeconomic variables change we need a) a multi party political competition and b) di erentiation among the public spending promises of the parties. Hence, we only require that simultaneously satis es both conditions. Luckily enough, we can guarantee that for almost all 15

18 values of. Of course, it might still be very interesting to study further those marginal cases, even if their analysis is not directly relevant for the purposes of studying electoral fragmentation. Nevertheless, we abstain from doing so here and we leave this task for future work. Finally, we want to stress that the equilibrium structure and hence, the outcome of the electoral game is robust to changes in the voting rule. To see this, recall that our preferences, within each group of voters, are exactly identical to those that Groseclose (2007) calls one-and-a-half dimensional preferences, where the rst dimension is the standard left-right axis and the second one is a good-bad dimension (high-low public spending in our case). In this second dimension voter s preferences within each group are identical. In our set-up all unemployed prefer high public spending (tax rate) whereas all the employed ones prefer low. Hence, we can invoke Groseclose s Theorem which extends Black s (1958) Median Voter Theorem to the one-and-a-half dimensional framework. Given that, in our set-up, both proportional representation and plurality rule pick the most preferred outcome for the median 26, it is clear that the outcome of the electoral game would have been identical under majority rule as well. 3 Empirical Analysis 3.1 Introduction In this section of the paper we bring our model to the data and we test our main result presented in Proposition 2: the non-monotonic relationship between unemployment and fragmentation. For this reason, we use aggregate political, socio-demographic and macroeconomic data we collected for twenty-three western OECD democracies during the period from 1960 to Bringing the model s predictions to data will be an interesting exercise for a variety of reasons. Firstly, as argued before and as the data in Fig. 2 and 3 suggest, our main prediction that at low and moderate levels an increase in the unemployment rate decreases the fragmentation of the party-system, might estrange some of our readers. Hence, the need for an empirical test of the Economic Voting hypothesis is indispensable. This will allow us to enrich the existing political cycles literature (Alesina, 1987) by endogenizing political parties preferences over unemployment. Moreover, so far there has never been a comprehensive and comparative empirical study on the economic determinants of voting behavior and electoral fragmentation that encompasses almost all western democracies. Thus, the present study is a rst attempt to document empirically the economic origins of party-system fragmentation in western democracies, not just incumbency in a speci c country. Of course, we stress once more that our current empirical study does not aspire to give a complete account on how and why citizens vote the way they do. Rather, we aspire to shed some light and provide some empirical evidence on the potential impact of economic variables, such as unemployment, on electoral fragmentation. Surely, unemployment is not the sole determinant of electoral fragmentation, yet we think that it is one of the most important ones. Finally, our theoretical model builds on a theory of rational voting based on real economic outcomes. Therefore, in order for the theory to stand scrutiny, it is important to put it through a reality check. Since all our key explanatory variables are aggregate macroeconomic data we think it is imperative to provide an empirical account on how well our theory ts the data. Hence, the basic estimation strategy is to conduct a cross-country panel data analysis in order to examine whether unemployment and electoral fragmentation are related in a non-monotonic fashion, just 26 In our case, the median is indi erent between parties L and R, which in equilibrium are elected with equal probability

19 as Fig. 8 suggests. Our initial hypothesis, summarized in Proposition 2, is that an increase in unemployment is associated with an decrease in the level of electoral fragmentation for values of unemployment less than 10 percent 27. For unemployment rates higher than this threshold value the direction of the relationship is reversed. The remainder of the empirical section is organized as follows: Section 3.2 provides an analytic description of our main sources of data. Then, in Section 3.3 we proceed by presenting our basic estimation strategy and model. Section 3.4 discusses some preliminary OLS results and potential sources of biases and endogeneity. In Section 3.5 we deal with problems arising from endogeneity by introducing our IV s and discussing the rationale behind their use. Finally, in Section 3.6 we present and discuss our main IV results and Section 3.7 concludes the empirical component of this paper. 3.2 Data Description To test our hypothesis, we have collected and compiled observations of key political, socio-demographic and macroeconomic variables from 23 western OECD countries. 28 We have decided to restrict attention to this group of countries for many reasons. Firstly, we need to focus on countries with fairly stable and fair democratic institutions and electoral procedures for the most part of the second half of the twentieth century. Secondly, given that democratic elections take place approximately every four years, we had to focus on countries which have a long history of democratic and fair elections. Adding other countries that are undergoing a democratization procedure would, in fact, reduce our time span and our number of observations. Finally, our key hypothesis, the unemployment paradox, would seem more plausible and to some extend expected in LDC s or developing economies that are under transition. Yet, we want to illustrate that even in mature and industrialized democracies economic voting still plays an important role in determining electoral outcomes. As a result, we decided to focus on OECD economies. In the remainder of this paper we will be using the observations of The Comparative Political Data Set which is a collection of political and institutional data which have been assembled in the context of the research projects Die Handlungsspielräume des Nationalstaates and Critical junctures:. An international comparison directed by Klaus Armingeon and funded by the Swiss National Science Foundation. It consists of a compilation of (mostly) annual data for 23 democratic OECD countries for the period of 1960 to 2007, taken from the SourceOECD Database. 29 We also supplemented the data set with observations collected from the OECD i- Library, for reasons of consistency, plus some own calculations and data from OPEC and the US Energy Information Administration (EIA). The data on electoral fragmentation from greek regional elections are own calculations based on the o cial results as they appear in the o cial database of the Greek Ministry of Interior. We have also supplemented our calculations with data from the Eurostat Regional Yearbook and OECD database for consistency. The data set is suited for cross national, longitudinal and pooled time series analyses. In what follows we organize this data set into a panel, in a way that is suitable for a cross-national pooled time series analysis. Although, as the empirical analysis proceeds, we will de ne in more detail some of the key variables which are used more extensively in our econometric analysis, for 27 See Fig. 7 where we calibrate the values of the parameters of our model using aggregate OECD data. 28 The term "western" refers to the those member-states of OECD that joined the organization prior to the collapse of the Soviet Union and "Warsaw Pact". 29 In the cases of Greece, Spain and Portugal, political data were collected only for the democratic periods. Data for Greece are missing during the period Data for Portugal are missing until 1975, and for Spain until

20 a complete description of the data set we refer the reader to the comprehensive list of all variables in the Appendix. Finally, the data set contains some additional demographic, socio- and economic variables. A few variables have been copied from a data set collected by E. Huber, Ch. Ragin, J. Stephens, D. Brady and J. Beck eld (2004), as well as from a data set collected by D. Quinn. For a more complete description of those variables we refer our readers to the OECD online database (OECD i-library). The data are summarized in Table 13 in the Appendix. 3.3 The Benchmark Model In deciding which estimation strategy to follow and how to choose the speci cation of our model, we had to take into account the limitations posed by the format of the data. Even though we have a total of 1,022 observations from 23 countries over a time span of almost fty years, our dependent variable, electoral fragmentation, exhibits low variation over time. The reason being, elections take place approximately every four years in almost all OECD democracies. With some notable exceptions, such as countries that have a "tradition" in calling frequent early elections, this is true for the majority of the countries in our sample. As a result, electoral fragmentation is a quite static variable. To deal with this complication we have decided to include fragmentation only on those years that elections took place. Hence, we have kept only the years that elections took place. Formally, following the literature we de ne electoral fragmentation as a Rae Index: F i;t = 1 NP i;t (v n;i;t ) 2 ; n=1 where v n;i;t is the vote share of party n in country i at election year t, and N i;t is the total number of parties in country i that contested elections in year t. In total, after eliminating non electoral years, we are left with 322 observations in all twentythree countries over a period of 48 years. This number of observation is completely within expectations, since it implies an average of 14 elections per country, with an election taking place on average every three and a half years. Secondly, we have decided to organize the data as a cross-country panel (pooled time series) in order to include country speci c xed e ects. The main idea behind this is the fact that electoral fragmentation might depend on time-invariant, country-speci c characteristics. As recent literature documents, electoral outcomes and hence, the fragmentation of the party system, can critically depend on some of those characteristics. These may include, for instance, the structure and the historical attributes of the political system, country speci c demographics, such as minority and ethnic parties, irregularities and disproportionalities in the electoral rules, urbanization, the structure of the political system itself, constitutional and institutional arrangements, or even the political history of each country, to name a few. Obviously, those parameters might pose a threat to the validity of our estimations if they are not taken into consideration. Hence, in order to account for all these factors, as well as others not explicitly named in this study, we use a xed e ects (FE) estimator after conducting the relevant Hausman test. Formally, we want to estimate the following model: 18

21 F i;t = q i;t + 2 q 2 i;t + 3 D i;t 1 + X 0 i;t + a i + t + i;t where the dependent variable F i;t is the modi ed Rae Index of electoral fragmentation measured in percentage terms only on election years, over the period , q i;t is the unemployment rate for country i in year t, D i;t 1 is the lag of the stock of public debt, Xi;t 0 is the set of exogenous covariates and i, t are country and year xed e ects. Our independent variable, unemployment, is measured as a percentage of the active labor force. We have included a quadratic term on unemployment to account for the non-monotonic relationship, suggested by the theory. Even though, as Fig. 8 suggests, almost 90% of our observations on q lie below the 10 percent threshold, making the documentation of the non-monotonic relationship very hard, a negative coe cient 1 and a positive 2 would speak in favour of our hypothesis. We have also decided to include the lag of public debt (measured as a percentage of GDP). The rationale for including it is related with our discussion on how the level of promised public spending can a ect the voting decision of the unemployed and hence, fragmentation. The higher the level of debt is, the lower the level of public spending that a government can promise is. As a result, the vote shares of dominant parties shrink and fragmentation rises. Since this positive relationship between debt and electoral fragmentation is supported by greek regional election data and, since it is also documented in the existing literature (Persson et. al, 2007), we decided to include the lag. The remaining controls include various measures of the disproportionality of the electoral rule (plurality vs. proportional rule) and electoral rule change, the degree of institutional constraints in abusing public nances (another key parameter that a ect public spending promises), the incumbency e ect, the type of government, the number of parties contesting the elections and voter turn-out. To control for some of those, we use a series of binary variables as exogenous covariates such as the existence of single-member district plurality rule and others. Yet, since it is virtually impossible to account for all possible political parameters that might have an e ect on fragmentation, we recognize that our OLS estimates might be biased towards zero, due to potential omitted variable bias (OVB). Later on, when we will introduce our Instrumental Variables, we will try to account for the bias caused by omitted exogenous covariates. But before doing so, we rst want to include in the model those covariates that are available and examine how they a ect the estimation results. Before presenting our OLS estimates, a small comment with respect to our decision to include time xed e ects is, we think, in order. We have stressed so far that the main aim of this study is to document the impact that unemployment has on voting behavior and electoral fragmentation. The optimal way to achieve this would have been to exploit year speci c, country invariant shocks on the economy. By doing so we could have been able to capture changes on unemployment that are beyond the control of domestic political competition. For instance, a global nancial crisis that started in Asia and then spread around Europe could have been an ideal source of variation for unemployment that cannot be attributed to the policies of domestic governments, not to mention the level of party-system fragmentation. Exploiting those exogenous variations on output, unemployment and public spending would have been ideal for the purposes of this study. Yet, a potential source of worry comes from the fact that voting behavior and as a result electoral fragmentation may also depend on year speci c political phenomena that a ect uniformly all western democracies. Such an example could be the fall of the Berlin Wall, or the 9/11 terrorist attacks, that had an impact on partisanship and voting 19

22 behavior (Kaplan and Mukand, 2011). Of course, the incidence of such political phenomena is more rare compared to that of macroeconomic shocks, which we ideally want to exploit. Nonetheless, we have decided to include year xed e ects under all our chosen estimation speci cations, in order to account for such phenomena, knowing that this choice will increase the di culty of our task. On the other hand, we wanted to fully insulate our model from this kind of critique, by making sure that our estimated model is correctly speci ed. 3.4 OLS Results We rst present our OLS results that will serve as a benchmark. The main coe cients of interest are 1 and 2. According to our theoretical model and the Economic Voting hypothesis we expect that 1 < 0 and 2 > 0. That is, we hypothesize that an increase in unemployment is associated with a decrease in electoral fragmentation. Yet, as the unemployment rate becomes larger and larger, the relationship is reversed. Table 1 presents our initial OLS estimates under ve alternative speci cations. Column (1) simply checks and veri es Duverger s Hypothesis that plurality and rst-past-thepost electoral rules are associated with bipartisan systems and hence, lower levels of electoral fragmentation. The coe cient on plurality ER is negative (ranging from 15:8 to 6:5) and always statistically signi cant at any conventional level. Moreover, the coe cient on institutional constraints is also positive, under all speci cations (from 2 to 0:5) which provides further evidence in support of our model. Recall that, by Proposition 3, the introduction of institutional (e.g. scal) constraints limits the ability of dominant parties to manipulate scal policies (e.g. increase public spending) in order to gain more votes (and decrease fragmentation). Hence, the positive relationship should have been anticipated. Moreover, the coe cient on incumbency is always negative (approximately 2:1) and statistically signi cant, as theory predicts. These ndings serve as a consistency check on the relevance of our data set. In Column (2) we add year xed e ects and some other control variables. Thus, it is normal to expect that the inclusion of year xed e ects will increase the bias towards zero. Nevertheless, as discussed in the previous section, we think that the correct speci cation of the model should include both country and year xed e ects. So, we have decided to include them in all subsequent speci cations of our econometric models (both OLS and 2SLS). Columns (3) to (5) are the ones that have the most interesting economic interpretation. First, note that the coe cient on unemployment 1 is always negative (from 0:12 to 0:9) and statistically signi cant in Columns 4 and 5, whereas 2 is always positive, just as predicted by the theory. Moreover, despite the small occurrence of high unemployment rates (see Fig. 9.a) 2 is surprisingly statistically signi cant under all possible speci cations. This is an initial piece of evidence on the correct speci cation of our model. Furthermore, note that as we add the quadratic term (in Columns 4 and 5) to account for the non-monotonic relationship predicted by theory, and the interaction term between unemployment and institutional constraints (Columns 3 and onwards) to account for the di erential impact of unemployment on electoral fragmentation in party-systems that constrain excessive public spending, our estimates on 1 increase both in magnitude (from :12 to :91) and also in statistical signi cance, even at the 5% level. Though far from being perfect, our estimates clearly con rm the key elements of our model. The results in Columns 4 and 5 show a clear and signi cant non-monotonic relationship between unemployment and party-system fragmentation. 20

23 Table 1: Basic OLS Model under Various Speci cations and Sample Sizes Dependent Variable: Rae Index of Electoral Fragmentation (%) Dependent Variable Full Restricted (Controls) Non US Electoral Fragmentation (%) OLS OLS OLS OLS OLS (1) (2) (3) (4) (5) Unemployment Rate (%) (0.154) (0.373) (0.45) Unemployment-squared (0.017) (0.021) Institutional Constraints* Unemployment Rate (0.296) (0.285) (0.306) First-Past-the-Post ER (Dummy) (0.778) (2.056) (2.882) (3.243) (1.271) Institutional Constraints (Dummy) (0.938) (0.886) (2.589) (2.582) (2.840) Incumbent Gov t. Defeated (0.627) (1.156) (1.069) (1.029) (1.120) L. Debt (% GDP) (0.02) (0.021) (0.021) (0.028) Country FE? Yes Yes Yes Yes Yes Year FE? No Yes Yes Yes Yes Const (4.663) (2.064) (3.033) (3.420) (2.569) Other Controls No Yes Yes Yes Yes Obs R F statistic Table 1: Note: Robust Standard Errors, clustered at the Country level reported in parantheses. (***) Statistically signi cant at the 1% level; (**) Statistically signi cant at the 5% level; Statistically signi cant at the 10% level. Sample size varies due to missing data on Debt for the period from In Column 5 we exclude the US from the sample. 21

24 In addition, the positive coe cient on the interaction term (0:28), one-third the magnitude of 1, shows that conditional on a country having high institutional constraints (e.g. inability to nance public spending through higher taxation or scal rules etc.) the e ect of unemployment on fragmentation is severely mitigated. This exactly con rms our results on the varying tolerance levels for unemployment among OECD countries (Corollary 2). Notice that when the quadratic term kicks in as well, the relationship is reversed and becomes positive. In our sample the turning point seems to be around 11 percent for countries with no scal or institutional constraints and around 8 percent for countries with high constraint. That is, in a country with high institutional constraints, when the unemployment rate approaches double digits, any further increase will cause electoral fragmentation to rise. If constraints were lower, this shift would have been delayed. Hence, as the level of constraints decreases and the ability to manipulate public spending increases, so does the tolerance level of unemployment. In the Appendix (Table 7), we present more OLS results under various alternative speci cations as a robustness check. We just note that we consider variables, such as voter turn-out and the dummy indicating coalition government, as outcomes that are either co-determined with fragmentation (e.g. turn-out) or determined afterwards (e.g. coalition government). For this reason, we will not use them in any future speci cation of the model. To the extend that someone argues, quite sensibly, that the political tradition and culture of a country that favours coalition governments (e.g. Germany or Netherlands) can have a positive e ect on electoral fragmentation, in contrast with single-party government tradition (e.g. USA), our answer is that since this tradition is a time invariant country-speci c characteristic of the respective political system and culture, which we believe it is the case, it is then captured by the inclusion of country xed e ects. Moreover, even if a country s tradition on having, or not, coalition governments is not time invariant, then according to Duverger (1954) and the vast strand of literature developed after his seminal work, the electoral rule crucially a ects electoral fragmentation. Hence, the probability of ending up with a coalition government is lower when Plurality electoral rules are applied. In fact, a simple check in our data set reveals a strong, negative and statistically signi cant correlation between plurality rule and incidence of coalition government ( 0:56). As a result, we decide to exclude those two variables from the list with the exogenous covariates, especially in light of the fact that we continue to include the dummy variable on the electoral rule, country and year xed e ects in all subsequent speci cations of the model. Finally, the main reason that our estimates on 1 and 2 appear to be biased towards zero is reverse causality. As we have stressed extensively in the introduction, there is a large strand in the literature that documents how policies and political outcomes can a ect unemployment and other economic variables. That is, we have reasons to hypothesize that not only unemployment has an e ect electoral fragmentation, but also at the same time electoral outcomes can a ect unemployment. Parties might pursue di erent policies with respect to unemployment. Therefore, apart from the standard OVB problem discussed earlier, that relates to the existence of some unobserved parameters that might a ect fragmentation, we also run into an endogeneity problem. With this in mind, in the next section we introduce oil price shocks as an instrument for unemployment. 3.5 Accounting for Endogeneity and Bias: Introducing Oil Price Shocks Although the OLS estimates seem to con rm our theoretical predictions, there are potential sources of worry. Therefore, we have decided to estimate the relationship between unemployment and electoral fragmentation using an IV estimator. The reason for choosing this speci c estimator are: 22

25 endogeneity and OVB. Since recent literature has documented that political cycles on the one hand (Barro and Alesina), and the fragmentation of the party system on the other (Persson, Tabellini et al.), can have an e ect on economic performance and macroeconomic variables, such as unemployment, it would have been di cult to estimate the direction of causality and get consistent and unbiased estimates with standard OLS models. To put it more simply, unemployment might be an endogenous variable, not only a ecting the outcome of the elections and the resulting fragmentation but also being determined by them. In order to deal with the endogeneity problem, we are instrumenting for unemployment, using the shocks on imported oil prices (at re nery) from 1960 and onwards. The idea of using the oil price shocks as an instrument for unemployment is rst discussed by Levitt (2001), in a paper where he summarizes the strategies for identifying the causal link between crime and unemployment. The di culty in identifying the causal e ect of unemployment on crime is similar with ours due to unobserved policies that a ect simultaneously both the crime rates and unemployment. Furthermore, reverse causality also makes its appearance. In a study at the US State level, Raphael and Winter-Ebmer (2000) test two instruments for unemployment: the closing of military bases and the shocks on oil price. They argue that oil price shocks can have an e ect on unemployment and thus, they decide to include it as an instrument. But apart from Raphael and Winter-Ebmer, we also have good reasons for including oil price shocks as an instrument for unemployment. A quick look at Panel 4, in the Appendix, reveals a certain pattern between oil prices an unemployment. Moreover, we argue that shocks on oil prices are not related neither with our exogenous covariates, such the electoral rule and the institutional constraints, nor with any other unobserved political variables that might a ect fragmentation. Since it is very hard to argue that the shocks on oil prices can cause changes in the electoral rule, or the structure of the party system in any given country, we think that it is an instrument that satis es the exclusion restriction. Of course, someone might argue that extremely high oil prices might cause public unrest and wide scale social disruption that might lead to a regime change or violent events. But, so far and to our knowledge, there is no evidence that oil prices can create such social dynamics which in turn might a ect the party system in a western industrialized country. This observation supplements our argument that the e ect of oil prices shocks on electoral fragmentation is solely coming through economic conditions and in particular unemployment. Furthermore, we argue that our instrument is independent of the outcome. That is, electoral fragmentation cannot have an e ect on the oil price shocks. On this latter point, there is a growing amount of empirical evidence that suggest that oil prices follow a pattern that is hardly a ected by the voting behavior in any OECD economy, not to mention by electoral fragmentation. Pindyck (1999) and Barnett and Vivanco (2003) show evidence on mean-reversion whereas Cashin, Liang and McDermott (2000) and Engel and Valdes (2000) nd evidence of persistence. Bartsch (2006) makes the point that the international oil prices show very weak mean reversion. Thus, we think that it would be extremely hard to argue that changes in electoral fragmentation in any OECD country can predict or a ect the shocks on the price of oil. This argument would amount to claiming that elections in a country can a ect the shocks in the price of oil. But, even in the extreme case that one is willing to claim that political outcomes in oil exporting countries can drive the shocks on the price of oil, this argument would not apply in our data set since we do not include major oil exporting countries, such as Russia or Saudi Arabia. The only oil exporting countries in our data set, which could potentially a ect the oil prices, are Norway and the US. 23

26 Yet, as a robustness check, we conduct our analysis in a restricted sample that excludes the oil exporting countries. Since we do not get signi cantly di erent results (see Table 6), we conclude that there is no reason for further worry. Using real imported oil prices as our exogenous instrument for unemployment, we want to estimate the following 2SLS model: q i;t = b 0 + b 1 P i;t 1 + Xi;t 0 + a i + t + " i;t and F i;t = ^q i;t + Xi;t 0 + a i + t + i;t, where P i;t 1 is the Producer Price Index-weighted (PPI) real price of imported crude oil at re nery and Xi;t 0 is the set of controls discussed in the previous section. One might question why we do not also include a quadratic term on unemployment. There are two reasons for not doing so. Firstly, it is extremely challenging to nd two valid instruments for the same variable. Instrumenting for both unemployment and its square would perhaps create problems of weak instrumentation. In Tables 2 and 3, we compare the OLS with our IV estimates using all possible sets of instruments and allowing for variable sample sizes. As expected, the OLS estimates were biased towards zero whereas under 2SLS the coe cient increases in magnitude and becomes a more negative number, as our theoretical model predicts. In fact, when we estimate the impact of unemployment on fragmentation using a 2SLS estimator, our coe cient estimates become statistically signi cant even at the 1% level. We attribute the improvement in our estimates in both the reduction of OVB, and the fact that by instrumenting unemployment with real oil prices we account for endogeneity. Moreover, comparing Columns (1) with (3) and (2) with (4) in both tables, we see that including the squared term makes the coe cients of interest become larger in magnitude. Furthermore, all tests on the validity of instruments (Hausman-Wu) and overidentifying restrictions (Sargan and Basman) fail to reject the null that our model is correctly speci ed at any conventional signi cance level. Then why do we decide not to include the squares? The answer is that in some speci cations our instruments are weaker than the standard convention. Contrary to that, when we do not include the square, our instruments are robust under any chosen speci cation, as Table 8 summarizes. Given that not including the squares runs against us, by decreasing the magnitude of the coe cients, we decide to strengthen our results by excluding it. Secondly, as shown in Fig. 9, almost 90% of our observations of unemployment rates are below the 10% threshold. Even if the theory predicts a non-monotonic relationship we would require more observations on the other tail of the distribution. Given that the majority of our observations concerns very small rates of unemployment, we can proceed by excluding from the sample countries like Spain and perform the analysis without the quadratic term. Here-in-after, we present our results on the restricted sample only (where we exclude the US and other oil-exporting countries) and we refer the reader to the Appendix for more detailed results. Since the choice of sample size does not a ect neither the sign nor the signi cance of our estimates, as Tables 2 and 3 reveal, we think that this is the optimal way to proceed for reasons of consistency, robustness and enhanced validity. 24

27 Table 2: Comparison of 2SLS and OLS Models under Alternative Speci cations on Full Sample Dependent Variable: Rae Index of Electoral fragmentation; Dependent Variable: Full Sample Electoral Fragmentation (%) OLS IV OLS IV (1) (2) (3) (4) Unemployment Rate (%) (0.1648) (0.5637) (0.3723) (1.4279) Unemployment-squared (0.0163) (0.1218) First-Past-the-Post ER (2.4495) (2.1824) (3.0809) (3.4194) Incumbent Govt. Defeated (1.0890) (0.8401) (1.0362) (1.6655) Dummy High Institutional constraints (0.2257) (0.3016) (0.2314) (0.1941) L.Debt/GDP (%) (0.0229) (0.0542) (0.0225) (0.0617) Const (3.2242) (2.3119) (3.6600) (5.4219) Country FE? Yes Yes Yes Yes Year FE? Yes Yes Yes Yes Obs Instrument Oil Price Oil Price Oil Price Oil Price R F statistic (excluded IV) Table 2: Clustered Standard Errors at the Country level reported in parantheses. (***) Statistically signi cant at the 1% level; (**) Statistically signi cant at the 5% level; Statistically signi cant at the 10% level. Of course, we recognize that none of the two solutions is ideal. But we also recognize that we need to nd a balance between consistency and validity. Using two instruments increases consistency but can lead to weaker instruments in some speci cations. On the other hand, not instrumenting for the square introduces a down-ward bias to our estimates. But, after comparing our estimates in Tables 2 and 3, it is clear that not including the square introduces a bias towards zero, both in the OLS and the IV estimators. Since the bias works against our claims, being able to nd statistically signi cant results under this speci cation further strengthens our point. Nevertheless, as a robustness check, we present our results with the inclusion of the quadratic term in the Appendix (Table 8). 25

28 Table 3: Comparison of 2SLS and OLS Models under Alternative Speci cations on the Restricted Sample (excluding oil producing countries and the US) Dependent Variable: Rae Index of Electoral fragmentation; Dependent Variable: Restricted Sample Electoral Fragmentation (%) OLS IV OLS IV IV (1) (2) (3) (4) (5) Unemployment Rate (%) (0.1877) (0.5513) (0.4506) (1.5398) (1.3857) Unemployment-squared (0.0201) (0.1798) (0.1454) First-Past-the-Post ER (Dummy) (3.4223) (2.0316) (4.5244) (8.1541) (6.0297) Incumbent Govt. Defeated (1.1762) (0.9834) (1.1122) (2.4888) (1.7362) High Constraints Interaction Dummy (8.1421) High Institut. Constraints (Dummy) (0.2377) (0.2885) (0.2421) ( ) (0.4661) L.Debt/GDP (%) (0.0294) (0.0601) (0.0301) (0.1048) (0.0762) Const (3.5799) (2.6333) (4.0373) (6.2296) (3.4775) Country FE? Yes Yes Yes Yes Yes Year FE? Yes Yes Yes Yes Yes Obs Sample Size Non US Non US Non US Non US Non US Instrument N/A Oil Price N/A Oil Price AR(2) R F statistic (excluded IV) Table 3: Clustered Standard Errors at the Country level reported in parantheses. (***) Statistically signi cant at the 1% level; (**) Statistically signi cant at the 5% level; Statistically signi cant at the 10% level. In column (4) we used three lags whereas, in (5) the model is just identi ed. In summary, our coe cient of interest 1, is expected to have a negative sign as predicted by the theory. Yet, we also expect that b 1 < 0, which means that in the e ect of an increase in the oil price is negative(positive) to unemployment(employment). Initially, our expectation on the sign of b 1 might sound counter-intuitive. Yet, it is consistent with a Phillips curve type of argument. Furthermore, by looking at the data in Panel 4 where we present a visualized version of the rst stage, we can get a rst indication on the expected sign of b 1. In fact, we will argue that, if our model is estimated correctly, b 1 must be small in magnitude, yet a negative number. Nevertheless, we defer further discussion on the economic relation of the rst stage regressions until we rst present our main empirical ndings in the section that follows. Then, we will go back one step, in order to justify our expectation on b 1. 26

29 3.6 Results of Instrumental Variable Regressions Table 4 below, presents our key estimates on the restricted sample of countries. Columns (4) and (6) in Table 4 present the estimates of the exact same model as in Columns (2) in Tables 3 and 2 respectively. The only di erence is in the size of our sample. In this table, with exception of Column (6) we restrict our attention only to those countries that are not producers of oil. We present our main estimates in Columns (4) and (5). A rst look reveals that in both cases the coe cient of interest on unemployment is negative and almost equal in magnitude, 1:53 in Column (4) compared to 1:8 in (5). Moreover, it is statistically signi cant even at the 1% level, under all possible speci cations. The negative and statistically signi cant coe cient on unemployment con rms our initial hypothesis that for relative low levels of unemployment an increase in unemployment rate causes a decrease in electoral fragmentation. In addition, the F statistic on the excluded instruments is above the desired threshold, which implies a statistically signi cant rst stage relationship which has a clear economic meaning (more on the next section). In this sense, it is reassuring to see that our model is very robust in this alteration of the sample size 30. The coe cient on unemployment can be read as saying that an one percentage point increase in unemployment can be associated with an 1.8 percentage points decrease in electoral fragmentation. Given that the average fragmentation level in the sample countries is 72 points, we can say that an 1% increase in the unemployment rate is associated with more than 2% decrease in electoral fragmentation. Or, to put it di erently, a variation of one standard deviation in unemployment (which is equal to 4 percentage points) can account for two-thirds of the variation in fragmentation, on average. For a sole real economic variable, we think that the e ect is quite impressive and yields support to our Economic Voting hypothesis. Furthermore, apart from our key explanatory variable, we note that a series of variables have signs in accordance with our predictions. Of special interest is the coe cient on the high institutional constraints interaction dummy which is positive (0:74) and statistically signi cant. It can be interpreted as saying that conditional on a country exhibiting high institutional constraints (e.g. scal rules, budget balance clauses etc.) that do not allow for higher taxes and high public spending, the impact of unemployment on fragmentation is mitigated by almost 40%. Exactly as predicted by our theoretical model and in accordance with our empirical ndings in Fig. 2, the positive coe cient captures the inverse relationship between public spending and electoral fragmentation. Adding to that, the positive coe cient on debt can be interpreted as an extra nancial constraint that operates in the same way with the institutional ones. Hence, an increase in debt implies less room for public spending and as a result higher fragmentation. Although it is not statistically signi cant, the sing of the coe cient is in line with previous results by Persson, Tabellini et al. (2007) 31. But, since this is not the focus of this study, we con ne ourselves in noting this positive, and to some extend expected, relationship between public debt and electoral fragmentation. 30 For more robustness checks on sample size, model speci cations and choice of instruments please refer to Table 6, in the next section. 31 Yet, the di erence between our estimates and theirs is that we hypothesize the direction of causality running the other way. 27

30 Table 4: Key IV Estimates of the Basic Model on a Restricted Sample (Dropped USA) Dependent Variable: Rae Index of Electoral fragmentation (%); Instrumental Variable: Lag of PP-Indexed Imported Oil Prices at Re nery Dependent Variable: 2SLS Electoral Fragmentation (%) Non-US Non-US Non-US Non-US Non-US Full (1) (2) (3) (4) (5) (6) Unemployment Rate (%) (0.6433) (0.6077) (0.6077) (0.5513) (0.6102) (0.5637) L. Debt/GDP (%) (0.0584) (0.0586) (0.0586) (0.0601) (0.0501) (0.0542) First-past-the-Post ER (2.2198) (2.0316) (3.0624) (2.1824) High Institut. Constraints (0.2885) (3.8492) (0.3016) High Constraints Interaction Dummy (0.3752) Govt. Change (1.1661) (1.1661) (0.9834) (1.0010) (0.8401) Country FE? Yes Yes Yes Yes Yes Yes Year FE? Yes Yes Yes Yes Yes Yes Const (1.7381) (2.3729) (2.6864) (2.6333) (3.8661) (2.3119) Obs Sample Oil Importers Oil Importers Oil Importers Oil Importers Oil Importers Full Instrument Lag Oil Price Lag Oil Price Lag Oil Price Lag Oil Price Lag Oil Price Lag Oil Price R F statistic (1 st stage) Table 4: Note: Robust Standard Errors, clustered at the Country level reported in parantheses. (***) Statistically signi cant at the 1% level; (**) Statistically signi cant at the 5% level; Statistically signi cant at the 10% level. Column (6) on Table 4, presents estimates on the complete sample of Countries (including USA and oil producers) as a reference point. Moreover, the signs of two other control variables are in line with the predictions of the literature. The coe cient on the electoral rule dummy is negative and highly statistically signi cant under all speci cations. This estimate reads as follows: the change from a pure PR-list system (D i = 0) to a rst-past-the-post, single-member district Plurality rule (D i = 1) is associated with a decrease in electoral fragmentation by more than 10 percentage points (30 if we consider the full sample where USA is included). This nding is in accordance with Duverger s Law that relates the fragmentation of the party system with the electoral rule. 28

31 Even more interesting is the sign on the coe cient on the dummy indicating a government change. Its value ( 2:5) is statistically signi cant and points out to a known and well documented relationship between unemployment, electoral fragmentation and incumbency. As the literature suggested, high unemployment reduces the incumbent s chances for re-election. Our results indicate that in elections where the incumbent is defeated fragmentation is lower on average. The explanation is that the opposition rises in order to defeat the government, resulting in less vote dispersion. Our economic voting hypothesis is saying something more broad than this. It says that when unemployment is relatively moderate and rising there is a transfer of votes from the smaller (extremist) parties to the dominant (moderate) ones due to rent-seeking by the unemployed. This transfer of votes, that causes fragmentation to decline, is a separate e ect from the transfer of votes between the two dominant parties that is picked up by the coe cient on the incumbency e ect. As a result, our estimated coe cient on unemployment ( 1:8) captures the net e ect on fragmentation after accounting for the incumbency e ect. The fact that the sign is always negative and statistically signi cant under all speci cations (even in OLS models) con rms empirically what has been known in the literature for quite some time The Economics Behind our Instrument: Oil Prices and Unemployment To understand the validity of our instrument we need to summarize the economic relationship between oil price shocks and unemployment. Firstly, there is a growing consensus among the literature, that the e ects of an oil price shock on unemployment were not constant overtime. Especially after the 1974 oil crisis, the impact of oil price shocks on the macroeconomy has changed. In a early paper, Sweder van Wijnbergen (1985) presents a theoretical model of inter-temporal analysis, where there is disequilibrium in the goods and labour market in the rst period, in order to study the e ect of oil price shocks on the labour market and the current account. Although his analysis focuses mainly on the oil crisis, it can provide some critical insight on what might have happened after the 1970s. He argues that in the very short run the supply shock in the goods market, caused by an increase in oil prices, dominates and causes unemployment to increase. Yet, in the medium and long run the labour market adjusts. The increase in the price level causes real wage to fall below the worker s productivity and hence rms are willing to employ more and produce more, given demand. Then, if the hike in in ation is met by a cut in the rate of interest, investment will increase and there will be excess demand in both labour and goods markets. Especially for the economies that run CA de cits the e ect of increased investment on employment will be larger 32. This increase in investment makes the long-run e ect of an increase in oil prices on employment to be positive. That is, the initial negative shock that drives unemployment upwards, is quickly reversed. Stylized facts on the US economy during the 1975 oil crisis con rm this analysis. Unemployment in the US reached its peak in mid-1975 to rebound shortly afterwards and start declining due to the e ect on the real wage. One potential explanation for this is the ability of enterprises and industry to better absorb the costs of higher oil prices by means of passing the e ect on the consumers or by having stocks of output so that it can suspend production or even shift production to less oil intensive goods. To the extend that consumers are more adversely a ected by rising oil prices compared to the industry, and to the extend that degree of substitutability is large enough, we can expect the e ect of falling real wage to dominate over rising costs of production. Hence, after an initial negative 32 We note that at the time of this analysis refers to (mid seventies) the developed European Union incorporated only 6 member states and many of our OECD economies (Greece, Spain, Portugal and others) were considered to be developing ones. 29

32 shock on employment, we can expect that soon enough employment picks up again. In support of this latter hypothesis comes the paper by Keane and Prasad (1996), who study the e ects of oil prices on real wage and employment, using panel data from the National Longitudinal Survey of Young Men in the US. In their paper, they nd that while the short-run e ect of an oil price increase on employment is negative, the long-run e ect is in fact positive. The main reason for this nding is that increases in the price of oil unambiguously cause real wage to decline at the aggregate level in all sectors of industry. Moreover, they estimate that the fall on real wage is on the level of 4-5% for a standard deviation increase in the price of oil (19% approximately). Since the e ect of an oil price increase on labour demand depends on the degree of substitutability between oil and labour, they explain their nding by noting that labour and oil are not gross but net substitutes. As a result, the fall in real wage is to be expected. Hence, they conclude that it is those changes in relative wages and the sectorial changes in the labour market induced by the higher oil prices that can explain the positive e ect on employment. Blanchard and Gali (2007) study the structural changes in the labour market induced by oil price changes, by applying structural VAR techniques in a sample of advanced OECD economies to estimate the impact of an oil price shock to unemployment, in ation and output. In this sense their ndings, given their time span and the set of countries are directly relevant to our present study. They derive three main conclusions: Firstly, the impact of oil price shocks on the macroeconomy has changed over time and in particular after the 1980s. In that respect the episodes of 1974 are not identical with those that followed afterwards. Secondly, they attest that oil price shocks overtime have driven labour markets to become more exible, and thirdly, that due the shocks the share of oil in production has decreased overtime. In particular, with respect to unemployment and real wage, they show that the magnitude and the direction vary between the episodes of 1973 and 1979 and the episodes in early While unemployment rises sharply during the episodes of the 1970s, this trend reverses in the episode where unemployment falls in a response to an increase in the oil prices. Moreover, in the episode unemployment is declining while oil prices are still on the rise. Given that our sample contains data after 1970 for the majority of the countries, it is normal to expect b 1 to be slightly negative. They estimated that CPI in ation and wage in ation reach their peak at Q.4 after the incidence of an oil episode. This implies that due to more exible labour markets and the fall in the real wage, employment starts to pick up during Q.4 and unemployment starts to decline in the same time. In addition, they show that this trend is more characteristic of the late 1990s oil shocks, especially for countries like Japan, France and Germany. In light of the above, our estimates on coe cient b 1 should not come as a surprise, given that we use the one year lagged PPI price of imported oil as our exogenous instrument. Table 5 below, summarizes the rst stage results of our 2SLS model, presented in Table 4. Columns (1) to (5) in Table 5 are the corresponding rst-stage regressions of the 2SLS models presented in Columns (2) through (6) in Table 4. We present more estimates on the rst-stage regressions for the full sample (including US) and other speci cations in Table 9, in the Appendix. Our estimates on the impact of an oil price increase on unemployment are robust under all speci cations. In all ve of them we nd the e ect to be negative and statistically signi cant at the 1% level. Furthermore, the F statistic on the excluded instrument is above the value of 10, under all alternative speci cations. Since the estimates on our rst stage regressions have an economic interpretation, and moreover they are highly statistically signi cant under all speci cations, we conclude that the lag or real oil prices is a valid instrument. 30

33 Table 5: First-stage OLS estimates of the Basic 2SLS Model on the Restricted Sample (Dropped USA) Dependent Variable: Unemployment Rate (%) Dependent Variable 1 st -Stage Regressions Unemployment Rate (%) OLS OLS OLS OLS OLS (1) (2) (3) (4) (5) L.Real Oil Price (at re nery) (0.0024) (0.0024) (0.0027) (0.0028) (0.0027) L.Debt/GDP (%) (0.0166) (0.0166) (0.0201) (0.012) (0.0127) Dummy First-past-the-Post ER (1.4456) (1.9446) (1.8417) (1.0371) High Institutional Constraints (0.1511) (1.4044) (1.0761) High Constraints Interaction Dummy (0.1484) Const (0.841) (0.8383) (1.5076) (1.5209) (0.8023) Country Fixed E ects Yes Yes Yes Yes Yes Year Fixed E ects Yes Yes Yes Yes Yes Sample Size US Dropped US Dropped US Dropped US Dropped Full Obs R F statistic Table 5: Note: Robust Standard Errors, clustered at the Country level reported in parantheses. (***) Statistically signi cant at the 1% level; (**) Statistically signi cant at the 5% level; Statistically signi cant at the 10% level. Column (5) is the same with column (6) on Table 4, on the complete sample of Countries (including USA) After conducting a series of tests, we conclude that our instrument is neither weak (as evidence from the rst stage show) nor invalid (as the Hausman test statistic suggests). In the Appendix, we reproduce the results of Table 4 under alternative speci cations using di erent versions of our instrument and other estimators (GMM, 3SLS). Since all of them are qualitatively equivalent, have the same sign across speci cations and moreover, they are numerically close, we have decided to focus only on this speci cation, since the instrument we use is the simplest possible one. 31

34 The reason for not employing more indexation in measuring oil price shocks is that we fear that introducing more complicated techniques will compromise our data and allow room for extra critique directed at our instruments. In a sense, if real oil prices are not a valid instrument for unemployment, then, it would be hard to argue that it is the instrument and not the chosen indexation or technique that caused this relationship to be statistically signi cant. But, as the reader can see in Table 10, even with the simplest possible measure of oil price shocks, the reduced form relationship con rms our intuition that oil prices are a valid instrument for unemployment. Furthermore, what is even more reassuring, is the fact that our 2SLS model is extremely robust to the use of alternative speci cations to measure the oil price shocks. In Table 12, we present the estimates of our Model in Table 4 using two alternative instruments: the de-trended component of the real oil price and the predicted residuals of an AR(2) process on the real oil price. In the rst case to de-trend the oil prices we use the Hoddrick-Prescott lter with smoothing parameter 6.25, given that we have yearly observations. The main reason for using the de-trended component of oil prices is the ability to capture the shocks on oil prices better. We present the estimates of this model on Column (5) on the restricted sample. Since the coe cient estimates are statistically signi cant at the 1% level, negative and almost equal in magnitude ( 1:7 compared to 1:5) we think that there is no need to elaborate further on the Hoddrick-Prescott technique. In the remaining columns, we use the predicted residuals on oil price, after we allow it follow an AR(2) process. There is extensive literature documenting that oil prices might follow a meanreverting process. Formally we estimated the following 3SLS model: P i;t = i;0 + i;1 P i;t 1 + i;2 P i;t 2 + u i;t so that the 2 nd Stage becomes: q i;t = b 0 + b 1^u i;t + X 0 i;t + a i + t + " i;t and the 3 rd Stage becomes: F i;t = ^q i;t + X 0 i;t + a i + t + i;t, where ^u i;t is the predicted residual of the above AR(2) process on P i;t and all the other controls remain the same as before. Once more, we observe that our estimation results do not change dramatically compared to our estimates in Table 4. All the coe cients on unemployment are highly statistically signi cant and the estimated values are all negative and quite close in magnitude, ranging from 1:41 to 1:78. Moreover, in all cases the F statistic on the excluded instrument is larger than the desired critical value. As a result, we have decided to include our estimates as a further robustness check, but we think that more discussion about the AR(2) process is not warranted in this section. We only commend that choice of an AR(2) process was far from being arbitrary. Given that we observed systematic and signi cant auto-correlation on all the residuals until the second lag, we concluded that an AR(2) process was the optimal way to get the most accurate estimations on the real oil price shocks. 32

35 Finally, to conclude this section, we include a nal table in the Appendix (Table 11), where a robustness check is performed with respect to the choice of our estimator. As we have argued in the beginning of this section, we think that the just identi ed 2SLS model is the optimal choice to predict the e ect of unemployment on fragmentation operating through oil price shocks. Yet, in Table 11 we estimate our benchmark model using the GMM estimator. Once more, the estimates on the coe cient 1 are negative and quite similar in magnitude, ranging from -1.5 to But this time, they marginally fail to be statistically signi cant at the 5% level. Another complication related with the GMM method is that we don t have enough clusters in order to construct clustered standard errors. For these reasons, we conclude that the 2SLS model is the optimal one. Regardless, our results are robust under both speci cations, which allows us to proceed in their causal interpretation. Table 6 above, summarizes those results. 33

36 3.6.2 A Structural Test on the Consistency of the Theoretical Model As the reader may recall, in Section 2.5 we have commented on the di erent values that the structural parameter, that measures the ideological distance between the two moderate parties, can take. There, we have shown that as! 0, the two extremists will be less willing to sacri ce even a tiny vote share, by promising lower public spending, in order to increase the probability of winning for their ideologically neighboring party, simply because it has converged to the median. Hence, they prefer to match-up the public spending promises of the moderate parties, leading to an equilibrium in which the public spending dimension is practically cancelled out since all parties propose identical scal policies in order to capture the vote of the unemployed. As a result, our proposed link, through which unemployment a ects electoral fragmentation, cannot be as e ective as before. Hence, if our model has correctly identi ed the economic voting mechanism through which unemployment a ects electoral fragmentation, we should expect that as moderate parties converge ideologically, fragmentation should not vary signi cantly with the unemployment rate. Or if it does so, the e ect has to be much more weak. The reason is that high ideological convergence will render our carrot-and-stick mechanism ine ective. As a result, since both moderate and extreme parties are o ering the same carrot, the voting behavior of the unemployed should not vary with public spending promises. This implies that if the our model is correctly speci ed, variations in unemployment should have a signi cantly smaller e ect on electoral fragmentation once the above channel has been shut o. In practical terms, this implies that if we want to bring the structural consistency of our model into a test, once we estimate our basic econometric model on a restricted sample of countries, where moderate parties exhibit a very high degree of ideological convergence, we should nd that the impact of unemployment on electoral fragmentation is much lower. Perhaps it might even fail to be statistically signi cant, as the coe cient 1 on unemployment will become a smaller in magnitude, yet negative number. To test this prediction, and of course to verify that our model is consistent, we split our sample into two separate groups. The rst group contains the countries where the moderate parties (those ones that take turns in government) have converged in terms of ideology, according to the Schmidt Index. The second one contains those countries where the dominant parties exhibit high ideological divergence. Figure 9.b shows the degree of ideological distance among dominant parties in the OECD countries from 1960 to Based on this separation, we estimate our 2SLS model on the rst group of countries. We present the results in Table 17, in the Appendix. It is reassuring to observe that the estimation results on the restricted sample of countries (sample size is reduced to half compared to our results in Tables 4 and 6) con rm our theoretical predictions. That is, they o er strong support in favor of our model s speci cations. We can see that in all speci cations, irrespective of the choice of our instrument and the inclusion or not of extra controls, the coe cient on unemployment is always negative (from almost 1 to 0:4) and strictly smaller in absolute value compared to the estimates in Table 4 ( 1:8). That is, on average, its size drops to almost a half. Yet, it fails to be statistically signi cant, as expected, under all alternative speci cations and for any conventional level of statistical signi cance. Also, note the fact that our rst-stage estimates are statistically signi cant at the 1% level under all speci cations, and also the fact that the coe cients on all other explanatory variables of interest (e.g. institutional constraints and incumbency e ects dummies) are statistically signi cant, have the same sign and are identical in magnitude with the respective ones in Table 4 (e.g. the coe cient on incumbency is around 2:1 whereas the coe cient on institutional constraints is around 0:7 in 34

37 both cases). Then, taking all these evidence into consideration, it becomes clear that our estimate on the coe cient for unemployment becomes smaller in magnitude and loses its signi cance due to the fact that we have restricted our sample in countries where the ideological convergence between moderate parties is very high. Hence, the data lend strong support for the model s economic voting hypothesis since the e ect of unemployment on electoral fragmentation runs through the channel that our model has speci ed. So, in summary, our empirical results yield enough support both to the model s structural elements and also to its predictions. 3.7 Discussion of Empirical Results The importance of our empirical ndings stems out from two factors. Firstly, our results yield support for the Economic Voting Hypothesis and its main underlying idea that dominant parties tend to take advantage of the relatively stronger rent-seeking voting behavior of the unemployed voters. So far, we have shown, by the means of theoretical modelling, the signi cant impact that the economy has on voting behavior, electoral outcomes and the structure of the party-system, as measured by the level of electoral fragmentation. Therefore, it is quite important to note that our empirical estimations yield extremely strong support to the ndings of the theory. In speci c, in our sample the average variation in the unemployment rate between elections is 4 percentage points, whereas the average variation in the level of electoral fragmentation is 9:5 percentage points. Given that one percentage point increase in unemployment is causing an almost 2 percentage-point decrease in electoral fragmentation, it is easy to see that one standard deviation in unemployment can account for more than 2=3 of one standard deviation in electoral fragmentation. Or equivalently, one standard deviation change in unemployment can be associated with a net 8 percentage-point change in the allocation of electoral power between parties, after accounting for the incumbency e ect. Even though we acknowledge that there are, of course, other factors, for instance the electoral rule that a ect fragmentation and might counteract with unemployment, thus, mitigating the variation in fragmentation, this is still an extremely astonishing number. Especially after taking into consideration the fact that we are only accounting for a sole economic variable. If we also account for the e ect of public spending (which we partially do through institutional constraints), then, we can get an idea of how important the role of the economy is in determining electoral behavior and party-system fragmentation. In fact, the economy seems to be the single most important determinant of electoral behavior in western OECD democracies. Secondly, apart from documenting the importance of economic conditions in determining electoral outcomes, the degree of its occurrence and the non-monotonic relationship between unemployment and electoral fragmentation allows us to revisit the traditional political business cycle theory. Our study empirically documents that the dominant parties exploit the fact that the voting behavior of the unemployed is mainly driven by rent-seeking. As a result, an increase in unemployment causes a reduction of electoral fragmentation. Hence, the dominant parties might bene t from higher unemployment, up to a certain point where non-monotonicity kicks in. And, although we do not go as far as to suggest that big parties have an opportunistic incentive to pursue higher unemployment, we note that our empirical ndings provide some reasoning as to why some parties (in our case the extremist ones) might favour scal discipline and the introduction of scal constraints (e.g. balanced budget amendments, scal rules etc.) over policies that aim to reduce unemployment. The latter argument brings into mind the current debate in the US for the need of adapting stricter scal rules and the role of the Tea-Party movement in leading this 35

38 discussion. This stands in direct contrast with common wisdom as expressed in the traditional opportunistic political business cycle theory, developed by Nordhaus (1975). Nordhaus model predicts that parties out of pure opportunistic motives would have an incentive to reduce unemployment before the elections. Yet, Alesina and Roubini nd that empirical evidence are not in accordance with Nordhaus opportunistic model predictions. Instead, they nd support for Alesina s partisan rational expectations model of political business cycles. But, Alesina s model starts from a point where parties have heterogenous preferences. Our model, in essence, bridges the gap between those two models. While it is a model of opportunistic parties, its predictions on the relationship between unemployment and the political cycle are more in line with the partisan rational expectations model. The main reason for this is the approach we follow, by reversing the link between political outcomes and the economy. In the Alesina model parties have exogenously given, heterogenous preferences on unemployment and in ation. In our model parties preferences on unemployment and debt are derived from the impact that economic conditions have on their chances for electoral success. That is, we endogenize parties preferences over economic outcomes and we explain how the data can t the opportunistic model. In that respect, even though we develop a model of opportunistic parties, our empirical results are similar with Alesina s partisan rational expectations model of political cycles. As a result, we o er a further insight on why parties might have di erential preferences over the economy. Therefore, our model is not antagonistic with theirs. Rather, it helps complete the political business cycle theory by exploring both directions of the link between partisan politics and the economy. 4 The Patron-Client Paradigm: Evidence from Greek Local Elections In the introduction we identi ed two potential links through which economic parameters, such as the unemployment rate, can a ect electoral fragmentation and the structure of the party-system. We have explored the rst one (economic voting) both theoretically and empirically, in Sections 2 and 3. Here, in this section, we exploit the action of the greek government to ddle the books and mis-report the level of public de cit for the period , which generated an information shock, as a natural experiment. Our aim is to demonstrate our secondary link, through which unemployment has an impact on electoral fragmentation: the existence of patron-client relations between politicians and voters that underpin the political system. Furthermore, we also want to test our secondary prediction with respect to the e ect of expected public spending on electoral fragmentation (Proposition 3). Using data from greek local elections is the best empirical strategy to follow, mainly for two reasons. Firstly, the greek political system is characterized by long-lasting clientelistic relations among politicians and voters. The origins of the clientelistic nature of the greek political system date back to the beginning of the twentieth century. Up until recently, when the independent Supreme State Council for Personnel Selection [Anotato Simvoulio Epilogis Prosopikou- ASEP] was instituted in 1994, the public sector was viewed as the prize for the election-winning party. Consequently, public sector appointments were primarily made according to partisan a liation and political favoritism, in complete absence of meritocracy 33. Furthermore, once the Council was instituted, the practice 33 Historically, the greek term used for describing this transaction was borrowed from the Ottoman word rüşvet [bribery], a clear analogy with the practice of buying public o ce followed by the Ottoman administration. 36

39 of non-meritocratic public sector appointments, and the subsequent clientelistic link, passed from the central government to regional administrations whose autonomy was instituted at the same time (1993) 34. Hence, partisan favoritism moved from elected parliamentarians to elected local administration o cers (e.g. mayors and regional governors). Secondly, and more importantly, the two step audit and revision of greek public nances that was published in October of 2009 and 2010 respectively, o ers the opportunity for a natural experiment. The reason is that while the rst revision was referring to the FY2009 public de cit, the second one was concerning the de cits accrued in the period from 2006 to That is, the second revision was a pure information shock that drastically changed the expectations of the public on the ability of the state and the local authorities to engage in public sector expansion and extensive public spending for many years to come. A brief look at Table 14 helps in making the point more clear. As we can see, the rst revision that took place in the period from May-September 2009, related to the FY2009 de cit. But, the second revision (May-September 2010) was mainly due to mis-reported de cits of the central government concealed by the greek authorities in the past. As a result, out of the 11.7 percentage points of the total revision almost all referred to the hidden de cits of the period. In fact, more than 80% of this revision was solely referring to the FY2006 de cit. Clearly then, this upward revision of public debt cannot be attributed to policies followed in 2009 or even Rather, it was already there but was not disclosed neither to the citizens nor the rest of the economic agents. That is why, we refer to this second revision as a pure information shock. The second increase in the level of public debt was, in fact, back-dated as it had occurred at least three years ago. What was indeed new is the revelation of this piece of information to the public that occurred just days before the November 2010 Local Elections. 4.1 The Information Shock as a Natural Experiment and the Timing of Events Before proceeding to the speci cs of our identi cation strategy and design, it is rst useful to present a brief time-line of events that lead to two consecutive revisions of greek public nances. Then, it will become evident why we can exploit this information shock as a quasi-experiment. The sequence of events is as follows: 1. In August of 2009, following its defeat on the elections for the European Parliament that took place in June, the conservative Greek Government [New Democracy Party - Nea Dimokratia] calls for early elections to take place in September. It cited its inability to execute the Budget for FY2009 as the main reason due to the fact that the forecasted 5% de cit for FY2009 had to be signi cantly revised upwards (it was projected to exceed 10%). 2. In October 2009, Eurostat in cooperation with greek statistical authorities announce the results of the audit and revise the FY2009 de cit upwards by almost 8% (approximately 12.8%). Elections take place right after the announcement, amidst a climate of tension caused by the incumbent government s misleading estimates on public de cit and the nancial audit process. As a result, the conservative government was largely defeated in the elections by the centre-left party [Greek Socialist Movement - PASOK ] 35. Yet, electoral fragmentation remained within historical high bounds (similar to 1996 when the two big parties collectively got the 79% of the vote). 34 In October 1994, the rst-ever elections for regional authorities (at the NUTS-3 level) took place in Greece. Of course, local elections for mayors and city-councilors were taking place since Yet, regional governance was instituted only in PASOK gained the 44% of the vote while Nea Dimokratia got a historically low 34%. 37

40 3. In April 2010, a new report published by Eurostat consolidates its estimates on the FY2009 de cit. According to these estimates the de cit soared at 15.9%, approximately 3 points more than anticipated. In the meantime the centre-left government-elect, in collaboration with EU agencies, announces a new scal audit that will cover all periods from 2005 and onwards. In May 2010 the greek government asks for the help of EU and IMF and a package of austerity measures is decided. 4. In October 2010, the nal report of Eurostat is published and the FY2009 de cit is nalized at 16%, almost 11% more than initially anticipated. Yet, in comparison with previous estimates, the total revision of the greek public debt for FY2009 alone is an astonishing 24% (from 102% to 126%). Almost half of it (11.7%) is attributed to hidden de cits in the period from when the previous government was mis-reporting the data. From this 11.7% more than three quarters (or 8.3%) was solely attributed to excess de cit in FY In November 2010, Regional elections take place, after citizens are exposed to this new information with respect to the hidden de cits throughout the period from 2006 to The incumbent government [PASOK] is not held accountable since it was the policies of the previous administration that lead to excess de cits and the subsequent austerity measures. As a result, it wins convincingly the local elections securing the victory of its preferred candidates in 8 out of 13 regional administrations (including the Athens Metropolitan Region with 3 million voters) and almost all major municipalities. Yet, as it can be seen in Figs. 3 and 10 electoral fragmentation rises and a new historical-high level is recorded. The main reason that allows us to exploit the information shock on the level of public debt comes from the fact that the new information was revealed prior to the local elections but after the national ones. Yet, even though public nances were already in bad shape since 2006, national elections took place without the voters knowing this information. While there is no actual change in public nances between October 2009 and October 2010 (there was a small update on the level de cit by 3% which was signaled by Eurostat earlier that year and, thus was anticipated), there is a shock on the information about the status of public nances in the past (there is an overall 12% revision on past de cits) which alters signi cantly voter s expectations on future public spending and the intensity of the required austerity and scal adjustment. Moreover, the fact that the change in expectations was caused by an information shock solves for us the apparent endogeneity problem, which we have extensively discussed in the previous section. If it were for current scal mismanagement causing the revision of de cit and debt in 2010, it would have been impossible to correctly identify the e ect of decreased public spending on electoral fragmentation because the causal link might have been reversed. But fortunately, in our case nothing has changed in the fundamentals of the economy between the two electoral races. The status of the economy was equally bad both in late 2009 and mid What changed, and with it so did the expectations of the public, was the information on how bad the scal position of Greece was back in But this cannot be attributed to current policies pursued by the current government. Rather, it is an outcome of past policies and most certainly of the clientelistic nature of the party-system in Greece. Hence, exploiting this information shock provides us with the ground for a natural experiment in order to test the impact of reduced expected public spending on electoral fragmentation through the link of clientelistic relations. 38

41 4.2 Data and Identi cation Strategy Before presenting our key ndings, in this section we will analyze our identi cation strategy in more detail. As we have argued, the information shock will allow us to capture the e ect of public spending on fragmentation. Yet, in order to identify whether the cliental nature of the greek political system has played any role in it we need to identify the regions were those relations are stronger and more persistent. Once we identify them, we should expect to nd a di erential e ect of the information shock on electoral fragmentation on those regions. In fact, this is what appears to be the case, if one looks at Fig. 10. Clearly the rise in fragmentation was larger in regions where public sector employment is higher, whereas in regions with low public sector employment the increase was not at all signi cant. The main identifying hypothesis is the following. The information shock altered voter s expectations on the states s ability to nance public spending and transfers to the voters and local authorities 36 for many years to come, especially after the IMF rescue plan. But this shock a ected disproportionately the regions that have a larger public sector share due to stronger and more resilient clientelistic relations 37. Therefore, in regions where public sector employment is larger, the expected cut in public spending will have a greater impact on people s expectations about future clientelistic transfers (e.g. government jobs) and hence, on their voting behavior. As a result, we should observe a larger increase in electoral fragmentation in those regions due to the fact that the clientelistic link is becoming less e ective. Formally, we want to estimate the following model: F i;t = 0 + s + t + t P 1 =t 4 D s; + t D s;t + X i;s;t + " i;s;t where F i;t = 1 NP i;t (v n;i;t ) 2 is the index of regional electoral fragmentation, de ned as before, n=1 N i;t is the number of parties in region i at time t, and v n;i;t is the vote share of a party n 2 N, s is the dummy indicating a High Public Sector Region ( s = 1, if regional public sector employment is higher than the national average), t is the post information shock time dummy ( t = 1, if t = 2010) and D s;t is the interaction ( High Public Sector)*(2010) Dummy. Finally, X i;s;t is the set of other controls (e.g. unemployment, local elections dummy, regional dummies etc.). Clearly the key coe cient of interest is t which captures the di erential impact of the shock on the high public sector regions. The basic unit of analysis is a NUTS-3 Region 38, of which Greece has 48. But we also repeat our estimates for the NUTS-2 level. Since we do not get dramatically di erent results, we decided to focus on the NUTS-3 regions in order to have a larger sample size. From those 22 NUTS-3 regions (or 6 NUTS-2 regions) that accounts for the 60% of the population are classi ed as High Public Sector Employment regions (quasi-treatment group), whereas the remaining 26 (or 7 at the NUTS-2 level) are classi ed as our quasi-control group. In those regions the main sources 36 Most of the funding in greek local administrations comes from the central budget and EU funds. Local authorities in Greece have extremely limited capacity to generate and raise revenue to nance their operations. Almost 80% of their funds come from the central government budget and EU structural funds. 37 Following the literature we use high public sector employment as a proxy for stronger clientelistic relations. 38 NUTS stands for Nomenclature Unitaire de Territoire Statistique and is the basic regional statistical unit of the EU. There are three levels of regional sub-units. At the NUTS-1 level Greece has 2 regions (North and South), at the NUTS-2 level it has 13 regions and 48 at the NUTS-3 level, the smallest possible sub-unit. 39

42 or economic activity and employment are non-government related (e.g. tourism and agriculture). Table 15, presents the summary of key statistics for both groups of regions. The data on regional electoral fragmentation were collected by the Greek Ministry of Interior and compiled by us. The nal calculations of electoral fragmentation indexes are our own. The time span of the data in the sample runs from 1998 until This includes 8 electoral races (4 at the local level and 4 at the national), each taking place every four years. We have also supplemented our data on electoral fragmentation with a series of economic and socio-demographic variables at the NUTS-3 level which we have copied from the Eurostat LFS Survey for reasons of consistency in our base sampling unit. In total we have 384 observations at the NUTS-3 level. Furthermore, in order to insulate our results we take three important steps: Firstly, we group the rebel candidates vote shares (expressing disagreement with their mother parties and in this particular case with the IMF memorandum) with the vote shares of the o cially endorsed candidates of their respective mother-parties. 39 But, we do so only for the post-shock electoral results and not for the pre-shock elections. Secondly, as discussed above we estimate our model using the NUTS-3 regions as our main unit of analysis. Going from the NUTS-2 to the NUTS- 3 level we secure that the metropolitan regions of Athens and Thessaloniki, the two areas that exhibit the highest volatility in fragmentation and the regions where the most sharp increase in fragmentation in the 2010 elections was recorded, are signi cantly under-represented in the sample. Rather than being two out of 13 regions (almost 16% of the sample) they are now recorded as two out of a total of 48 regions (only 4% of the sample). Moreover, in some speci cations we completely exclude those regions to account for the education e ect. That is, since the level of college graduates is higher in those regions than any other region, and given that literature suggests that more educated voters are more likely to perceive and process faster new information with regards to the economic situation it might be the case that the di erence in fragmentation we pick up is driven by this e ect. Yet, by excluding those regions completely we can account for this. Since our results do not change dramatically, we mention this as a robustness check. Finally, we use pre-shock lags (the terms t P 1 D s; in the estimated equation) in order to analyze the trend of the treatment e ect over time. That is, there might have been recorded in the past a wide di erence in electoral fragmentation level among regions with high and low public sector employment. So, we might be picking up this trend rather than the impact of the information shock that generated expectations for lower public spending. Nevertheless, the inclusion of lags can account for this successfully. And as our results indicate in the next section, the di erence we pick up is attributed to the information shock. The general idea of all these steps is to arti cially suppress downwards the post-shock levels of electoral fragmentation for the quasi-treatment regions (in crude numbers the di erence is cut down in half). If we still manage to get a statistically signi cant and large in magnitude positive coe cient on the interaction term D s;t then, this would imply that the e ect we are capturing is real. Before presenting our main results, and in order to address some concerns that might arise due to the fact that we are using observations both from national and regional elections, we need to make a brief comment on the nature of electoral competition in Greece both at a local and national 39 Greek Local elections, as evidence in Figs. 11, 12 and 13 show, have a avour of national mid-term elections. As a result all main parties that are represented in the Parliament o cially endorse candidates for each regional electoral competition both at the NUTS-2 and -3 levels. Moreover, in most of the cases those candidates are high-pro le party members. Party members who disagree with those choices or want to express a general disagreement against the pursued policies of party leadership may enter the race as rebel candidates. This phenomenon is very common in greek local elections since their institution. =t 4 40

43 level. This is the topic of the next section. Here, we brie y refer to the graphs in Figs. 12 and 13, where we compare the trend of electoral fragmentation between national and local elections both within and across groups of regions. As we can see, there seems to be no signi cant di erence in the trend either within or across groups. Since the parallel trends assumptions is satis ed, we can proceed with our estimation. Nevertheless, in all estimated speci cations we use a local elections dummy as a control variable to account for any unobserved di erential trend, not appearing in the eye Greek Local Elections: A Primer on clientelism Greek local elections at the NUTS-3 level (regions) were rst introduced in 1994 and take place regularly every 4 years. The electoral rule is two-round majority. Hence, there is no lost vote argument in the rst round since voting for a non-winning candidate endorsed by a smaller party won t a ect the nal outcome 40. That is, there is no incentive for strategic voting in the rst round. That is why we measure electoral fragmentation during the rst round when all candidates and parties participate in the electoral contest. Moreover, the two-round majority rule with a run-o ensures that no coalition administration ever takes place. In, fact this has never occurred so far. Therefore, there is no need to worry about the impact of the electoral rule on fragmentation through coalition administrations. Political parties represented in the Parliament always endorse openly candidates for local elections. Partisan support and a liation is extremely important for election (only a couple of NUTS-3 administrations had an independent and non-a liated chair-person). Local administration, especially at the NUTS-3 level, has extremely limited revenue raising capabilities. Almost 80% of its revenues comes from direct central government subsidies and EU structural and peripheral funds. Hence, local administrators acts as decentralized distributors of government money. As a result, clientelism and partisan favoritism are very strong in local elections, taking the most usual form of public sector appointments. Furthermore, Greece has a stable bipartisan political system since 1977 with no coalition governments ever occurring (with a small six month exception in the summer of 1989). That is, the Persson-Tabellini argument that stresses the link between coalition governments and high public spending is not applicable in our case. Moreover, electoral fragmentation ranged historically between 60-65%, meaning that the greek party-system is characterized by stable bipartisanship. Clientelistic voter-party relations are the epicenter of political competition, since for years the two major parties used their in uence in the administration to favour their clientele (public sector appointments and other privileges). The number of parties contesting elections and represented in Parliament remained relatively stable (3-4 minor parties in addition to the two major ones never exceeding 20% of the vote). The two major parties (PASOK and Nea Dimokratia) are getting around 80% of the votes. Local elections are highly politicized, taking place on the year between national elections, in analogy with the US Congressional ones that take place in the year between two successive presidential elections). All greek parties represented in the Parliament, always endorse openly candidates for local elections and partisan a liation is a very strong predictor of electoral success. Hence, electoral and voting behavior in local elections is almost identical with national ones. As 40 If there is a winner in the rst round this means that she got more than 50% plus one of the total votes cast. Hence, voting strategically would have no e ect. On the other hand, if no candidate gets the absolute majority in the rst round, the run-o takes place among the two rst ones. 41

44 a veri cation test, Figs. 11, 12 and 13 provide a very good visual argument. Both within and across groups, voting behavior and electoral fragmentation in local elections appears to follow an identical trend, in most cases coinciding with the outcomes of national elections. Therefore, we conclude that local elections act as a proxy for mid-term national Elections. Since the political characteristics between national and local elections are identical we need not worry in our analysis. Nevertheless, as a consistency check, we do include a dummy variable for local elections. 4.3 Estimation Results: NUTS-3 Level In Table 16, in the Appendix, we present the results of our main estimates under many alternative speci cations. In all of them our key estimate, the coe cient on the interaction dummy between high public sector regions and the post-shock time variable is positive, large in magnitude (from 2:7 to 4:5) and statistically signi cant at any conventional level. Especially in Columns (4) and (5), when we include the lags, the coe cient almost doubles in magnitude. The positive value of t con rms our hypothesis that the expectation of reduced public spending (and less rents) has led to an increase in electoral fragmentation in regions with high public sector employment and stronger cliental relations. The point estimate on the coe cient can be interpreted as saying that an almost 12% increase in the level of de cit that caused public spending expectations to adjust is associated with an increase in electoral fragmentation of 4.5 percentage points more in the quasi-treatment regions. This corresponds to almost an 8% increase in the average level of fragmentation. Hence, we conclude that the e ect is also economically signi cant. Moreover, the coe cient on unemployment is negative ( :15) and statistically signi cant at the 5% level. This is also in line with our theory that predicts a negative relationship between unemployment and electoral fragmentation for moderate levels of unemployment. Furthermore, what is interesting is the positive coe cient (0:3) on the interaction term between unemployment and high public sector dummy, double the value of the coe cient on unemployment. Both of them together can be interpreted as follows: the expected cut in public spending that generated expectations for lower rents and government jobs caused a di erential response among the unemployed. In regions where cliental relations are stronger, the expectation is that they are going to be more adversely impacted by this cut in government spending and clientelistic job creation. As a result, the initial negative relationship is completely reversed. This nding comes as further evidence to the support of the hypothesized non-monotonic relationship between unemployment and electoral fragmentation. In this case though, the link is di erent. Now, an increase in unemployed caused an increase in fragmentation because there is no expectation of job creation or rent distribution by the government since the cliental basis of the political system is severely undermined (public spending becomes unavailable). But this e ect is only observed in those speci c regions, that used to exhibit stronger clientelistic links (high public sector employment). In all other ones the e ect is still negative. Collecting together all the evidence, we conclude that the expected reduction in public spending had an undisputed positive e ect on electoral fragmentation via the link of cliental relations. Hence, our empirical ndings con rm once more the predictions of our model with respect to the negative relationship between public spending and electoral fragmentation. 42

45 4.4 Discussion of the Results Our case study on greek local elections provided empirical support for the patron-client link through which economic conditions and especially public spending a ect electoral fragmentation. More speci cally, through a natural experiment that exploits an information shock, we showed a causal relationship between limited expected public spending and increased party-system fragmentation. We have also highlighted the importance of constraints ( scal or institutional) in reshaping the political landscape and the structure of the party-system. Hence, we have argued that scal discipline, austerity measures and the reforms in public sector, which were adapted by the greek Parliament in order to tackle the excessive public de cit, acted as an exogenous constraint to the amount of disposable public spending. This constraint facilitated and accelerated the dismantling of the clientelistic basis of the greek party-system. Although we do not go as far as claiming that the generalized public spending cuts signaled the end of bipartisanship in the greek political system, we can say with certainty that they have weakened the resilience and strength of clientelistic relations. This, in turn, undermined the dominant position that the two major parties enjoyed, undermining the bipartisan foundations of the party-system which were laid on the patron-client relation between voters and politicians. As a result, the scal constraints imposed by the austerity reforms liberated involuntarily the voters (clients) from their patrons (dominant parties) and increased the level of electoral fragmentation and political pluralism. Yet, this increase in the vote shares of smaller parties came at the cost of both an increased level of polarization and extremism in public a airs and also drastically reduced political participation and voter turn-out, as the data suggest. Overall, it is too early to conclude whether the bene ts of increased political pluralism and representation and weaker clientelism will outrun the potential costs of reduced stability and moderation that the bipartisan system was o ering. Nevertheless, no matter which side of the argument proves to be valid, the key point remains: the impact of economic parameters, such as public spending and unemployment, and of clientelism on the structure of the party-system is of outmost signi cance. 5 Concluding remarks To sum up, in this paper we have attempted to give a more comprehensive explanation to the phenomenon of varying electoral fragmentation in western OECD democracies, taking into account the impact of economic factors on electoral behavior. Rather than arguing in the traditional way, that the degree of electoral fragmentation depends solely upon the existing political and institutional arrangements, we have presented a theoretical model of electoral competition that relates electoral fragmentation with economic conditions. In speci c, we have shown how unemployment a ects electoral fragmentation via two links: opportunistic (economic) voting and clientelism. The two major contributions of our work are the formal formulation and the empirical veri- cation of the "Economic Voting" hypothesis, which highlights the rent-seeking elements in the voting behavior of the unemployed voters, and the resulting non-monotonic relationship between unemployment and electoral fragmentation (negative for low values of unemployment). Of course, this is not to say that opportunistic, rent-seeking voting behavior is the sole determinant of electoral fragmentation in a political system. Certainly, there are other factors, such as ideology or institutional design that a ect citizens voting decisions, which in turn determine the level of fragmentation. Yet, as our theoretical model suggests, there is a signi cant relationship between 43

46 economic conditions, such as unemployment in the rst place and public spending, and fragmentation. Additionally, the model also o ers and interesting insight, based on rational choice theory, as to why extremist parties target the more ideologically driven voters and lobby for more policy constraints (e.g. balanced budget rules, no taxation to nance public spending etc.) while the centrist ones might focus more on tangible political promises, such as increased public spending. Apart from the theoretical predictions, our empirical ndings yield strong support to the predictions of the model and document the impact of the macro-economy on electoral fragmentation. In that respect, our study is pioneering, since it is the rst to document systematically the economic origins of party-system structure and the relationship between unemployment and electoral fragmentation. In fact, we nd that this relationship is quantitatively important since variation in unemployment accounts for 2/3 of the variation in net fragmentation, controlling for the incumbency e ect that economic conditions have. Moreover, it appears to be extremely robust to variations of the model. Furthermore, by exploiting the information shock on Greek public nances we are able to document an alternative link through which our e ect operates that of clientelistic relations among parties and voters. We also document the impact that public spending expectations have on electoral fragmentation. Hence we cover both sides of the "carrot-and-stick" story. Taken altogether, our paper highlights the importance of economic conditions as determinants of party-system structure and electoral behavior. Moreover, the empirical ndings in support of our hypothesis allow us to study the link between economic conditions and electoral outcomes in both directions. Thus, we contribute in enriching and completing the theory of political business cycles. And even though our paper does not explicitly provide a theoretical model of political cycles, it clearly hints the intuition behind such a model. Since some political parties, driven by opportunistic motives, might prefer to promote policies that stress the importance of scal discipline at the expense of unemployment reduction, then, our empirical ndings provide the ground to reconcile the stylized facts of "partisan rational expectations" political cycles theory with the theory of the traditional "opportunistic" one. Yet, we leave its formal exploration for future research. Finally, on a more technical scope, our paper provides further support to the existing literature for using real oil price shocks as an instrument for unemployment. Given that there is a limited amount of studies which document the use of this instrument, our results add further justi cation to this growing body of empirical literature. Moreover, although not the main purpose of this study, our empirical results con rm the ndings of other studies about the di erential impact of oil price shocks on employment both between the short-run and the long-run and also between the late 1970s and the early 1990s. 44

47 References [1] Acemoglu, Daron, S. Johnson, and J. A. Robinson (2005), Institutions as a Fundamental Cause of Long Run Growth, Handbook of Economic Growth, Vol.1, Part 1: [2] Acemoglu, Daron, and J. A. Robinson (2006), Economic Origins of Dictatorship and Democracy, Cambridge University Press, Cambridge and New York. [3] Acemoglu, Daron, and J. A. Robinson (2000), Why did the West Extend the Franchise? Democracy, Inequality, and Growth in Historical Perspective, Quarterly Journal of Economics, Vol. 115(4): [4] Alesina, A., "Macroeconomic Policy in a Two-Party System as a Repeated Game" (1987), Quarterly Journal of Economics, Vol. 120(3): [5] Alesina, A. and N. Roubini, "Political Cycles in OECD Economies" (1992), Review of Economic Studies, Vol. 59(4): [6] Arcelus, F., and A. H. Meltzer, (1975). The e ects of Aggregate Economic Variables on Congressional Elections. American Political Science Review 69: [7] Armingeon Klaus, Panajotis Potolidis, Marlène Gerber, and Philipp Leimgruber (2009). Comparative Political Data Set I Institute of Political Science. University of Berne. [8] Barro, Robert J (1996). Institutions and Growth, an Introductory Essay, Journal of Economic Growth 1(2): pp [9] Bartsch, Ulrich (2006), How much is Enough? Monte Carlo Simulations of an Oil Stabilization Fund for Nigeria, IMF Working Paper 06/142 (Washington: International Monetary Fund). [10] Blanchard, Olivier J. and J. Gali (2007). The Macroeconomic E ects of Oil Shocks. Why are the 2000s so Di erent from the 1970s?, NBER Working Paper No [11] Bloom, Harold, and D. Price (1975). Voter Response to Short-run Economic Conditions: The Asymmetric E ect of Prosperity and Recession. American Political Science Review 69: [12] Campbell, James E., and James Garand (2000). Before the Vote: Forecasting American National Elections. Thousand Oaks, CA: Sage Publications. [13] Cashin, Paul, Hong Liang and C. John McDermott (2000), How Persistent Are Shocks to World Commodity Prices?. IMF Sta Papers 47(2): [14] Colomer J.M. (2005). It s Parties that Choose Electoral Systems (or, Duverger s Laws Upside Down), Political Studies, vol. 53: pp [15] Cox, Gary W., and J. Morgan Kousser (1981). Turnout and Rural Corruption: New York as a Test Case, American Journal of Political Science 25(4): [16] Dixit, Avinash and J. Londregan (1996). The Determinants of Success of Special Interests in Redistributive Politics. The Journal of Politics, vol. 58:

48 [17] Duverger, Maurice. Political Parties: Their Organization and Activity in the Modern State, Wiley Press, New York, [18] Engel, Eduardo and Rodrigo Valdes (2000), Optimal Fiscal Strategy for Mineral Exporting Countries. IMF Working Paper 00/118. (Washington: International Monetary Fund). [19] Esteban, Joan, and Debraj Ray (2011). Linking Con ict to Inequality and Polarization, American Economic Review 101(4): [20] Dziubiński, Marcin, and J. Roy (2011). Electoral Competition in 2-Dimensional Ideology Space with Unidimensional Commitment. Social Choice and Welfare 36: [21] Ergun S.J., (2010). From Plurality Rule to Proportional Representation, Economics of Governance, Vol. 11(4):. [22] Fiorina, Morris. (1981). Retrospective Voting in American National Elections. New Haven, CT: Yale University Press. [23] Groseclose, T. (2007), One and a Half Dimensional Preferences and Majority Rule. Social Choice and Welfare 28: [24] Holbrook, Thomas M. (2008). Incumbency, National Conditions, and the 2008 Presidential Election. PS: Political Science and Politics, 41: [25] Kaplan, Ethan and S. Mukand, (2011). The Persistence of Political Partisanship: Evidence from 9/11, mimeo [26] Keane, Michael P., and E.S. Prasad (1996), The Employment and Wage E ects of Oil Price Changes: A Sectoral Analysis, Review of Economics and Statistics, Vol. 78(3): [27] Kiewiet, D. Roderick (1983). Macroeconomics and Micropolitics: The Electoral E ects of Economic Issues. Chicago: University of Chicago Press. [28] Kinder, Donald, Gordon S. Adams, and PaulW. Gronke (1989). Economics and Politics in the 1984 American Presidential Election. American Journal of Political Science 33: [29] Kinder, Donald, and D. Roderick Kiewiet (1979). Economic Discontent and Political Behavior: The Role of Personal Grievances and Collective Economic Judgments on Congressional Voting. American Journal of Political Science 23: [30] Kinder, Donald, and D. Roderick Kiewiet, (1981). Sociotropic Voting: The American Case. British Journal of Political Science 11: [31] Kramer, Gerald, (1971). Short-Term Fluctuations in U.S. Voting Behavior, American Political Science Review 65: [32] Krasa, Stefan, and M. Polborn, (2010). The Binary Policy Model, Journal of Economic Theory, 145(2): [33] Laakso, Markku and Rein Taagepera (1979). E ective number of Parties: A measure with Application to West Europe. Comparative Political Studies 12, No.1:3-27. [34] Levitt, Steven D. (2001). Alternative Strategies for Indentifying the Link Between Unemployment and Crime, Journal of Quantitative Criminology, Vol. 17(4):

49 [35] Meltzer, Allan H., and S. F. Richard (1981). A Rational Theory of the Size of Government, Journal of Political Economy, Vol. 89(5): [36] Nordhaus, William D., (1975). The Political Business Cycle, The Review of Economic Studies, Vol. 44(2): [37] Persson T. and G. Tabellini (2003), The Economic E ects of Constitutions, MIT Press, Cambridge, MA. [38] Persson T., G. Roland, and G. Tabellini, (2007). Electoral Rules and Government Spending in Parliamentary Democracies, Quarterly Journal of Political Science, Vol. (): [39] Pyndick, Robert (1999). The Long-run Evolution of Energy Prices. Energy Journal, Vol. 20(2): 127 [40] Rae, Douglas W. (1968). A note on the fragmentation of Some European Party Systems, Comparative Political Studies 1: [41] Reny, Philip J., (1999). On the Existence of Pure and Mixed Strategy Nash Equilibria in Discontinuous Games, Econometrica 67(5): pp [42] Shapiro Carl, and J. E. Stiglitz, (1984). Equilibrium Unemployment as a Worker Discipline Device. The American Economic Review 74(3): [43] Tufte, Edward (1978). Political Control of the Economy. Princeton, NJ: Princeton University Press. [44] Van Wijnbergen, S. (1985). Oil Price Shocks, Unemployment, Investment and the Current Account: An Intertemporal Disequilibrium Analysis, Review of Economic Studies. 52(4):

50 6 Appendix A: Proofs Proof of Lemma 1. Continuity and di erentiability of z(:) are derived straightforward from the assumed properties of f(:). Then, for (i) consider the function z(q; m; ) and de ne: m f(:) f(m + ) f(m) and M f(:) f(m) f(m ). Then, z(:) takes the following form: z(:) = q m f(:) (1 q) M f(:) Taking a rst-order Taylor expansion we have: m f(:) f(m + ) f(m) = f 0 (m) and M f(:) f(m) f(m ) = f 0 (M ). We need only show that for every concave f(:) the following is true: q m f(:) > (1 q) M f(:), or qf 0 (m) > (1 q)f 0 (M ), which implies f 0 (m) > f 0 (M ). For any concave f(:) this, in turn, implies: m < M q(m m) =) q(m m) < M m =) q < 1 which is always true 8q; 2 (0; 1). Hence, z(:) is always positive. For (ii) rst note that m + (:) < M (:) =) m + (1 q)(m m) < M q(m m) =) (M m) < M m which is true for all 2 (0; 1). Then, with some abuse of notation, we can rewrite (q) and (q) as functions of q. Then, z(:) becomes: z(q; :) = q[f(m + (q)) f(m)] (1 q)[f(m) f(m (q))]. Then, we 2 z(q;:) = 2 f 0 (m + + f 0 (M @q 2 q f 00 (m + (q)) + (1 q) f Observe as @ 2 z(q;:) = [f 0 (M (q)) f 0 (m So we can write the 2 [qf 00 (m + (q)) + (1 q)f 00 (M (q))] 48

51 We need to show [qf 00 (m + (q)) + (1 q)f 00 (M (q))] < 0. < 0 for all f(:) and 8q; m; 2 (0; 1). Since f 00 (:) < 0 we have that: 2 < 0 i f 0 (M (q)) f 0 (m + (q)) < 0 () f 0 (M (q)) < f 0 (m + (q)) () M (q) > m + (q) which is always true. Hence, z(q; :) is concave w.r.t. q. For (iii) again with some extra abuse of notation, we can write () and () as a function of. Then, = qf 0 (m (1 q)f 0 Now, we need only > 0. But, = (1 q)(m = q(m m). That = > 0 [f 0 (m + ()) f 0 (M ())] > 0 () f 0 (m + ()) > f 0 (M ()) () m + () < M () which is always true for all q; m; 2 (0; 1). This completes the proof. Proof of Proposition 1. In order to show that the proposed equilibrium is indeed a symmetric NE, we need to show that no party has an incentive to deviate unilaterally from its equilibrium strategy, that is: 8p 2 P; V p (t p; t p) > V p (t 0 p; t p); 8 t 0 p; t p. First, we calculate the vote share v p (t p ; t p ) that each party receives as a function of its strategy (tax rate proposal) t p, for every t p ; t p 2 [0; ]. To do so, we have to identify the voter who is indi erent between voting for party r or R, R or L and L or l respectively. Then, we can compute the vote share for each party. We begin by identifying the indi erent voter between parties r and R 41. Formally, the following equality must hold for an unemployed voter (y = m): U y;x ( 1 2 ; t r) = U y;x (; t R ) =) j 1 2 xj + f(m + T (t r )) = j xj + f(m + T (t R )) () ( 1 2 x) + f(m + (1 q)(m m)t r ) = (x ) + f(m + (1 q)(m m)t R ) () x = [f(m + (t R)) f(m + (t r ))] 42 By analogy, for an employed voter (y = M) the condition becomes: 41 We need only examine the case where x 2 [; 1 2 ], since by single-peakedness and the fact that in equilibrium party r o ers t r = 0, any voters to the left of will never face the dillemma between L and r for that matter. 42 Here with some slight abuse of notation we de ne (t p) (1 q)(m m)t p and by analogy, (t p) q(m m)t p. 49

52 x = [f(m (t 2 r)) f(m (t R ))] Given that a fraction q of the electorate has y = m and the remaining 1 q has y = M, and given that the two continua of voters are identical in all other respects, we can then compute the aggregate indi erent voter: x = fq[f(m + (t R)) f(m + (t r ))] (1 q)[f(m (t r )) f(m (t R ))]g: Then, all voters to the right of x will vote for party r whereas all voters to the left of x (and till voter ) will voter for party R. By a symmetric argument a similar analysis applies when we compare the voter who is indi erent between parties L and l. We now repeat the same exercise for parties R and L 43. In a symmetric equilibrium by de nition t L = t R. Hence, the following equality must hold: U y;x (t L ) = U y;x (t R ) =) j xj + f(y + T (t L )) = j xj + f(y + T (t R )) Since T (t L ) = T (t R ), we have that f(y +T (t L )) = f(y +T (t R )) for every y 2 fm; Mg.Hence, we conclude that the indi erent voter is the median (x = 0). Then, we can compute the (symmetric) vote share allocation for each party as a function of its strategy choice (by symmetry it su ces to do so for parties R and r): v R (t R ) = fq[f(m + (t R)) f(m + (t r ))] (1 q)[f(m (t r )) f(m (t R ))]g and v r (t r ) = fq[f(m + (t 2 2 r)) f(m + (t R ))] (1 q)[f(m (t R )) f(m (t r ))]g Then observe that we have: h R = 1 ) R f 0 (m + (t R )) (1 R f 0 (M (t R )) h r = 1 r f 0 (m + (t r )) r f 0 (M (t r )) and, Clearly, since = @t we p > 0 i f 0 (m + (t p )) > f 0 (M (t p )) which by Lemma 1 is always true 8q; m and all t p 2 [0; ] such that 2 (0; 1). Hence, the vote share v p for every party p 2 P is strictly increasing in t p. The following Lemma summarizes a key feature of every symmetric allocation. Lemma 2 Let ^ 1 2 max q;m; hz(q; m; )i. Then for all > ^ and every t R, strategy t R = (and t L = respectively) is strictly dominant for party R (and L respectively). Proof. To prove the above statement we will evaluate how strategy t R = does against the worst possible case when t l;r = and t L;R = 0. Then, by an analogous computation we compute that: 43 We focus on the voters on the interval [ ; ] since the other voters are not relevant for this comparison. 50

53 v R (t R = 0) = z(q; m; ) and v r(t r = ) = z(q; m; ) =) v l = v r > v R = v L As a result, the two extremist parties win with equal probability. But consider now the following deviation: t 0 R = : Then, the resulting vote shares are: v R (t 0 R = ) = z(q; m; ) and v l = z(q; m; ) =) v 2 2 R > v p ; 8p 2 P r R Hence, party R now wins with probability one. Since its vote share is strictly increasing in the tax-rate, any proposal of t R 6= will do strictly worse in terms of vote-share allocation and outcome against any other strategy chosen by its opponents. Hence, strategy t R;L = is strictly dominant for parties R and L. This completes the proof of Lemma 2. Now, by Lemma 2 we can iteratively eliminate all other strategies for the two moderate parties and conclude that in equilibrium it must be the case that both R and L play t L = t R =. But then, the two extremist parties can never win in equilibrium. As a result they face a constrained vote maximization program. Taking t L = t R = as given, assume for a moment that p > 0 they also choose t 0 r= t 0 l =. By Lemma 2, in equilibrium, parties L and R win with equal probability. Then, any of the extremist parties (in our case r) has an incentive to undercut l (that is to propose t 00 r < t 0 l ) and cause party R to win with certainty i : v r (t 0 r = ) v r (t 00 r) The above expression just says that party r should o set the net loss in votes from undercutting with the net gain from causing party R to win and implemented its ideology. The net bene t from this change in outcome is equal to. But given t R = t L =, we know that the maximum loss in votes is equal to the following expression: max hv r (t 0 r = ) v r (t 00 r = 0)i = 1 z(q;m;) q [f(m + ) f(m)] (1 q) [f (M) f(m )] 2 2 8t r ; t r But then notice that we have de ned ^ 1 max hz(q; m; )i 1, which in turn implies that 2 q;m; 2 for every > ^ undercutting is always pro table. 44 That is, the constraint is always satis ed with strict inequality. Formally, this means that: > ^ 1 2 max q;m; hz(q; m; )i = max hv r(t 0 r = ) v r (t 00 r = 0)i Since by symmetry, undercutting is always pro table for party l as well, we conclude that in equilibrium we must have t r = t l = min ht j t 2 [0; ]i = 0. Hence we conclude that for every > ^ the unique Nash equilibrium strategy pro le is t = (t l ; t L ; t R ; t r) = (0; ; ; 0). This completes the proof. 44 By Lemma 1 we have shown that z(:) is a positive, continuous and strictly concave function. It can be checked that the expression z(:) q[f(m + ) f(m)] (1 q)[f(m) f(m )] obtains its maximal when q = 3, m = 0 and = 1. Hence, it is bounded above 4 away from 1 and the existence of an 2 (0; 1 ) such that > ^ is guaranteed. 2 51

54 Proof of Theorem. The second part of the Theorem is just a rephrasing of Proposition 1. It su ces to consider the degenerate mix-strategies l (t l = 0) = r(t r = 0) = 1 and L (t L = ) = R (t R = ) = 1. Then, clearly = E[ l (t l)] = E[ r(t r )] = 0. For < ^, note that by Lemma 2 L and R have a strictly dominant strategy t L = t R =. Then, w.l.o.g. we can x t L = t R = and restrict attention to the reduced two-player game G = (T p ; V p ) p2fl;rg, where T p = [0; ]. Then, we have a standard game with discontinuous payo s. Observe that for < ^ the constraint v r (t 0 r = ) v r (t 00 r) is not always satis ed (see Fig. 15.b). As a result, undercutting is not always pro table. That is, there exist values of such that the gain in implemented ideology () does not su ce to o set the incurred loss in the vote share. Hence, the game has no pure-strategy equilibria 45. Yet, there exist a symmetric equilibrium in mixed strategies. First, note that game G = (T p ; V p ) p2fl;rg is a symmetric in pure strategies, compact, Hausdor game, since T p is a compact Hausdor space. In order to show that game G possesses a symmetric mixed strategy Nash equilibrium we need only show that its mixed extension G is better-reply secure 46 along the diagonal, since quasi-symmetry of G follows from the symmetry of G (Corollary 5.3; Reny, 1999). Consider the mixed extension of the game G = ( p ; V p ) p2fl;rg, where we extend each V p to = l r by de ning V p ( l ; r ) = R 0 R 0 V p(t l ; t r ):d l d r for all ( l ; r ) 2. Then, in turn, better reply security of G implies two conditions: (i) reciprocal upper semicontinuity and (ii) payo security along the diagonal. For (i) we only need to verify that the sum of the payo s of the two parties P p V p(t) is u.s.c. in t on T. Then, by Proposition 5.1 (Reny, 1999) P R p V T p(t):d is also u.s.c. in on and the mixed extension game G is reciprocally u.s.c. The payo function for party r (and by symmetry l) is as follows (see also Fig.16): 8 < V r (t r ; t l ) = : v r(t r ; t l ), if t r < t l v r(t r ; t l ), if t r = t l v r (t r ; t l ), if t r > t l 9 = ; Then, by continuity of v p (t p ; t p ) for all t p ; t p and 8p 2 fl; rg, it is clear that P p V p(t) = 1 + P p v p(t) is continuous in t on T. As a result, condition (i) is trivially satis ed. For diagonal payo security in mixed strategies we need to show that: 8p; 8" > 0; 8 2, 9 ^ p 2 p s.t. V p (^ p ; 0 p) V p () ", 8 0 p in some open neighborhood of p. Clearly, this is always true. To see this pick any pro le = ( l ; r ) such that l = r and consider party r playing strategy ^ r that assigns larger probability to t r = 0 such that ^ r (t r = 0) > r (t r = 0). Then, for small perturbations of 0 l, close enough to l, the condition is always satis ed since there is at most a small loss in expected vote share that can be o set by a positive change in the expected outcome. As a result, since both conditions are satis ed we 45 Clearly, t r = t l = cannot be an equilibrium because a Bertrand-style induced competition will eventually lead to t r = t l = 0. But this cannot be an equilibrium either because r (or l) can go all the way and promise t r = since the gain in vote share max v l (t 0 l = ) v l(t 00 l = 0) = ^ exceeds the loss in utility that is now incurred by the fact that party L wins with certainty. 46 We would like to thank Philip Reny for suggesting the use of better-reply security in prooving the result. 52

55 conclude that mixed extension game G is better-reply secure. Hence, 8 < ^ the reduced game G possesses a symmetric Nash equilibrium in mixed strategies (t) = ( l (t l) = r(t r ); L (t L) = R (t R)), such that l and r have nite support on [0; ] with E[ l (t l)] = E[ r(t r )] = <, whereas L, R are the degenerate strategies with L () = R () = 1. This completes the proof. Proof of Proposition 2. Recall that we have computed: F (q; m; ) = 1 2C z(q; :) [z(q; :) + 2]. Then, we @z(:) = 2[z(q; m; ) By Lemma 1, we have shown that z(q; :) is a positive, continuous in [0; 1], di erentiable and strictly concave in (0; 1) function w.r.t. q. Hence, since also > 0, we have that 2[z(q; :) + ] < 0 for all q 2 (0; 1) and 8 m; 2 (0; 1). Then, we need only > 0 for all q 2 (0; ~q) and 8 m; 2 (0; < 0 for all q 2 (~q; 1) and 8 m; 2 (0; 1). Since z(q; :) is continuous on [0; 1] and strictly concave and di erentiable on (0; 1) we can apply Rolle s version of the Mean Value Theorem. Observe that: z(q = 0) = 0 and z(q = 1) = 0. Then, by Rolle s Theorem we have that: 9 ~q 2 (0; 1) such that z 0 (~q; :) = 0 But notice that z 0 (~q; :) = 0 also implies that F 0 (~q; :) = 0. Moreover, by strict concavity of z(q; :), we also have that ~q is unique. Hence, we have > 0 and consequently < all q 2 (0; (q;:) < 0 and consequently > 0 for all q 2 (~q; 1). That is, F (q; :) decreasing for every q < ~q and strictly increasing for every q > ~q. This completes the proof. Proof of Corollary 2. Recall that we have @z(:) = 2[z(q; m; ) But then, = f(m+(q)) f(m)+qf +f(m) f(m (q)) (1 q)f 0 : By a First-Order Taylor expansion = f 0 (m) + qf 0 (m + + f 0 (M (q)) (1 q)f 0 (M or 53

56 = f 0 (m) f 0 (m + (q)) + f 0 (M (q)) f 0 (M (q)) = = [f 0 (m) f 0 (M (q))] [f 0 (m + (q)) f 0 (M (q))] Further note that z 0 (~q)= 0 in turn implies that: (~q)[f 0 (m) f 0 (M (~q))] = (~q)[f 0 (m + (~q)) f 0 (M (~q))] But since m < m + < M for all q; m; 2 (0; 1) we have that: f 0 (m) f 0 (M (~q)) > f 0 (m + (~q)) f 0 (M (~q)) > 0. As a result: (~q) (~q) = 1 ~q ~q = f 0 (m+(~q)) f 0 (M (~q)) f 0 (m) f 0 (M (~q)) 1. Express the root ~q that solves z 0 (~q; :) = 0 as a function of m and, ~q(m; ) such that z 0 (~q(m; )) = 0. Then, observe that the above expression implies that ~q 1. As m increases, 2 observe that the ratio converges monotonically to 1. That is f lim 0 (m+(~q)) f 0 (M (~q)) = 1. m!m f 0 (m) f 0 (M (~q)) Hence, ~q(m; )! ~q(m;). As a result we have that < 0. Furthermore, as converges to 1 the numerator f 0 (m + (~q)) f 0 (M (~q)) monotonically converges to 0 and as a result so does the ratio. Hence, this implies that ~q(m; )! 1. That is, we ~q(m;) > 0. This completes proof. Proof of Proposition 3. We have to show < 0. @z(:) = 2[z(q; m; ) Since we know that 2[z(q; m; ) + ] < 0 and by Lemma 1 we have shown > (;:) conclude that < 0 for all 2 (0; 1) and every q; m 2 (0; 1). This completes both parts the proof. 7 Appendix B 7.1 Graphs and Figures 54

57 55

58 56

59 Figure 1: Fractionalization Levels in 23 OECD Countries from

60 Fig. 2: Electoral Fragmentation (Rae Index in %) and Unemployment Rates (%) in OECD Economies from (national averages) Fig. 3: The Information Shock on De cit and its e ect on Public Spending and Electoral Fragmentation 58

61 Panel 4: Oil Price Shocks and Unemployment in the OECD ( ): First stage relationship Year Log Real Oil Prices Unemployment (%) Fig.1: USA Year Log Real Oil Prices Unemployment (%) Fig.2: United Kingdom Year Log Real Oil Prices Unemployment (%) Fig. 3: Switzerland Year Log Real Oil Prices Unemployment (%) Fig. 4: Sweden Year Log Real Oil Prices Unemployment (%) Fig. 5: Spain Year Log Real Oil Prices Unemployment (%) Fig. 6: Portugal Year Log Real Oil Prices Unemployment (%) Fig. 7: Norway Year Log Real Oil Prices Unemployment (%) Fig. 8: New Zealand 59

62 Year Log Real Oil Prices Unemployment (%) Fig. 9: Netherlands Year Log Real Oil Prices Unemployment (%) Fig. 10: Luxemburg Year Log Real Oil Prices Unemployment (%) Fig. 11: Japan Year Log Real Oil Prices Unemployment (%) Fig. 12: Italy Year Log Real Oil Prices Unemployment (%) Fig. 13: Ireland Year Log Real Oil Prices Unemployment (%) Fig. 14: Iceland Year Log Real Oil Prices Unemployment (%) Year Log Real Oil Prices Unemployment (%) Fig. 15: Greece Fig. 16: Germany 60

63 Year Log Real Oil Prices Unemployment (%) Fig.17: France Year Log Real Oil Prices Unemployment (%) Fig.18: Finland Year Log Real Oil Prices Unemployment (%) Year Log Real Oil Prices Unemployment (%) Fig. 19: Denmark Fig. 20: Canada Year Log Real Oil Prices Unemployment (%) Fig. 21: Belgium Year Log Real Oil Prices Unemployment (%) Fig. 22: Austria Year Log Real Oil Prices Unemployment (%) Fig. 23: Australia 61

64 Fig. 5a: Electoral Fragmentation for calibrated values of m and (OECD averages) Fig. 5b: Electoral Fragmentation for m = 0 and = 1. 62

65 Fig. 6a: Tolerance levels of Unemployment ~q(m; ), as a function of m; 2 [0; 1]: Fig. 6b: Graphical Representation of Propositions 3 and 4 63

66 Fig. 7.a: Net gain in votes for a party proposing high tax-rate as a function of q and Fig. 7.b: Net gain in votes for a party proposing high tax-rate as a function of q and m 64