Satisfaction with Democracy and the Winner-Loser Debate: the role of policy preferences and past experience

WWW.SOCPOL.UNIMI.IT Dipartimento di Studi Sociali e Politici Università degli Studi di Milano Working Paper 03/2010 Satisfaction with Democracy and the Winner-Loser Debate: the role of policy preferences and past experience WWW.SOCPOL.UNIMI.IT Dipartimento di Studi Sociali e Politici Facoltà di Scienze Politiche, via Conservatorio 7-20122 Milano - Italy Tel.: 02 503 21201 02 503 21220 Fax: 02 503 21240 E-mail: dssp@unimi.it Luigi Curini (Università degli Studi di Milano) Martedì 16 marzo 2010 - ore 12,30 Aula seminari

Satisfaction with Democracy and the Winner-Loser Debate: the role of policy preferences and past experience Luigi Curini, Università degli Studi di Milano Willy Jou, University of California, Irvine Vincenzo Memoli, Università del Molise March draft Not to be quoted without permission of the authors Abstract Previous studies have examined the gap between electoral winners and losers on various dimensions of political support, from specific institutional evaluations to more diffuse aspects concerning democratic principles and procedures, and found that winners express significantly greater support than losers at each level. However, these works define winners and losers in a static fashion by examining their attitudes at only a single time point, and employ a dichotomous categorization that neglect possible variations within each group. In this study we take into consideration both the past history of winning or losing on present attitudes, and the impact of ideological distance from the government on a widely used indicator of system support, i.e. satisfaction with democracy. By analyzing surveys covering 31 countries, we show that the relationship between winner/loser status and satisfaction with democracy contains both a marginal dynamic nature as well as policy content. Winning for the first time always boosts satisfaction with democracy. Moreover, the impact of winning status increases as voters perceive the cabinet as close to their own ideological position. For repeated winners, however, satisfaction only increases if they are ideologically close to the government; otherwise their winning status has no effect. Two normative conclusions can be derived: first, governments that, ceteris paribus, minimize the ideological distance between itself and the largest number of citizens (including both winners and losers), may serve to increase political support. Second, cabinet alternation helps to boost the overall level of satisfaction with democracy by expanding the proportion of voters who have experienced winning at least once in recent elections. Keywords: satisfaction with democracy, election winner and loser, proximity theory, multilevel model

Satisfaction with Democracy and the Winner-Loser Debate: the role of policy preferences and past experience Introduction As a mechanism ensuring accountability of the governors to the governed, elections constitute one of the defining features of democracy. Election results shape government composition and policy outputs, most likely to the advantage of voters who support the winning parties. Previous studies have addressed the gap between electoral winners and losers on various dimensions of political support, from specific institutional evaluations to more diffuse aspects concerning democratic principles and procedures (e.g. Anderson and Tverdova 2001; Anderson et al. 2005; Moehler 2009), and found that winners express significantly greater support than losers at each level. However, these works define winners and losers in a static fashion by examining their attitudes at only a single time point, and employ a dichotomous categorization that do not capture possible variations within each group. In this study we take into consideration both the past history of winning or losing on present attitudes, and the impact of ideological distance from the government on a commonly used indicator of political support, namely satisfaction with democracy. Since winners stand behind parties or candidates currently holding the reins of power, it would come as no surprise that they express greater satisfaction in political authorities and institutions. As a corollary, one expects their satisfaction to decline if their parties lose power. In contrast, in mature democracies, winners have no stronger claim as owners of the political system than losers, since by definition democratic elections permit the fortunes of both current winning and losing parties to fluctuate at the ballot box, and thus offer losers the possibility of winning in the future. It follows that a distinction exists not only between present winners and losers, but between voters who have previous (especially recent) experience of winning and those who do not. The former should express greater approval in the political system even if their parties are currently in opposition, since the system has worked to their benefit, whereas the latter have little cause for satisfaction because their parties have long been excluded from power. In short, our study takes a dynamic view in defining winners and losers, taking into account not only present status but also past experience, and demonstrates that both current winners who had lost in the past and current losers who had won in the past are more satisfied with democracy than voters whose parties have not won before. In addition to a static definition of winners and losers, most extant literature also treat both categories as homogenous entities, an approach that neglects probable variations within each group. Assuming policy as the primary factor influencing satisfaction levels, we argue that not all winners can equally expect to see their preferred policies implemented, and some losers may derive greater policy benefits from the government than others. For both winners and losers, levels of satisfaction depend on how closely the government s preferences align with their own. For example, under a socialist-communist coalition government, socialist winners may see more of their goals turned into policy outputs than communist winners, and social democratic losers who are closer to the government may have less cause for dissatisfaction than conservative losers. In short, our study takes a step beyond the simple dichotomous categorization of winners and losers to consider the policy distance between individual 1

voters and the government mean, and shows that ideological proximity exerts a significant impact on satisfaction with democracy. The study is organized as follows. In the following two sections, we review the literature on political support and the winner-loser gap, and present the two central arguments of our paper, the policy performance and dynamic effect hypotheses. We then introduce data and measurements, including details on individual and aggregate level control variables. Next, we present and analyze findings on both hypotheses, and conclude by discussing the implications of our results. 1. Conceptualizing democratic support Easton (1965, 1975) describes the citizens evaluation of the institutions (performance) and its components (parties and political actors) as well as the citizens identification with the State as the fundamental elements of political support. He also distinguishes among three political objects - the political community, the regime, and the authorities in terms of diffuse and specific support. The former refers to loyalty to one s own political community, while the latter is based on the fulfilment of demands or satisfaction with outputs. Whereas diffuse support is more stable, specific support is susceptible to fluctuation depending on the performance of political institutions and actors. Norris (1999) and Dalton (2004) refine the Estonian framework by distinguishing five levels of political support. Specifically, they take into consideration three different objects at the regime level: regime principles express the normative values of the political system, reflected in the belief that democracy is the best form of government (Dalton 1999); regime performance refers to democratic processes (see Norris 1999); finally, regime institutions denote confidence in political institutions. Previous studies found that winners are more trusting of institutions such as the electoral commission, courts, the police (Moehler 2009), and the bureaucracy (Anderson and Tverdova 2001). Concerning performance, winners are more inclined to believe that government can be trusted to do the right thing and carry out its functions in the interests of all citizens (Anderson and LoTempio 2002). Finally, winners generally offer a more positive assessment of how democracy works (Banducci and Karp 2003; Cho and Bratton 2006). One should also note that losers consent toward principles underlying the regime, or regime-based trust (Nadeau and Blais 1993) arguably affects the long-term viability of democracy. In this study we use satisfaction with democracy as an indicator at the regime performance level of political support. According to Norris (2006: 6), satisfaction with democracy is indeed an indicator of public evaluations of how well autocratic or democratic governments work in practice. Although satisfaction with democracy is one of the most commonly utilized indicators of political support, its interpretation has not always been clear. On the one hand, some scholars claim its use is complicated (Linde and Ekman 2003) and its meaning appears somewhat ambiguous (Norris 1999). On the other hand, different indicators have been used by various authors in order to operationalize this concept (see Castillo 2006; Klingemann 1999; Dalton 1999; Kornberg and Clarke 1994; Montero and Gunther 1994). These debates notwithstanding, democratic satisfaction should be considered an expression of short-term evaluation of system outputs (Waldron-Moore 1999) and a useful measure of the discrepancy between democratic norms and actual processes (Wagner et al. 2003). 2

2. Exploring the Winner/Loser Effect (and its political implications) In the past decade a large number of works have analyzed the determinants of satisfaction with democracy, using both individual and aggregate data (see Norris 1999, Dalton 2004). One factor that has received increasing attention is how individuals who support parties in government (winners) and those who support parties in opposition (losers) differ in support for democratic processes and outputs. Although many studies have stressed the declining role of political parties at least since the end of the 1960s, parties still appear to have the capacity to influence citizens political views (Dennis and Owen 2001) and determine 'who gets what, when, and how' in a given polity (Lasswell 1953). In this sense, if citizens decide to vote for party A, it is because they believe it can better respond to their demands, satisfy their goals, and reflect their values. Since voters prefer victory over defeat, election outcome may affect attitudes toward not only the political system, but also the democratic process in general. When a voter chooses a party (or a candidate) that wins the election or at least enters into a coalition government, she expresses greater confidence in the government (Anderson and Tverdova 2001), which translates into increased satisfaction with democracy as well (Anderson and LoTempio 2002). Exactly the opposite tendency characterizes election losers, who are almost always significantly less satisfied with the way democracy works than winners (Anderson and Guillory 1997; Anderson and Tverdova 2001). 1 Losers discontent may affect the quality, even stability, of democracy. Hence Anderson et al. refer to losers as the crucial veto players of democratic governance (2005:7), and Nadeau and Blais also argue that the viability of electoral democracy depends on its ability to secure the support of a substantial proportion of individuals who are displeased with the outcome of an election (1993:553). The winner-loser distinction should be consequential in new democracies, where democratic rules of the game may not be well entrenched, and where some voters socialized under former authoritarian regimes who are not (yet) accustomed to the concept of losing may seek non-democratic means of pursuing their political objectives. There also exists the risk that winners would be too compliant toward the government, to the extent of being willing to overlook violations of democratic principles (Moheler 2009). However, the relationship between winner/loser status and satisfaction with democracy is neither immediately self-evident nor straightforward. To examine this causal linkage in more detail, we focus on two different aspects. a. The Policy Performance Effect First, previous research portrays voters assessments of electoral results as basically a zero-sum game (as noted in Henderson 2004). In other words, it is hypothesized that those who backed winning candidates will be more satisfied than those who backed losing candidates regardless of other conditions. Yet we know that even parties and voters on the same side of the winner-loser dichotomy can hold different ideological positions, and this could clearly matter for political satisfaction. According to the large body of literature on the spatial theory of voting (see Downs 1 One should note that governments can control legislative majorities based on a minority of votes. This means that losers can outnumber winners among the electorate. 3

1957; Adams, Merrill and Grofman 2006), voters are assumed to vote for the party presenting the closest policy program to their policy preference (see however Kedar 2005). Indeed, voters are assumed to derive utility from the package of policies implemented by the future government as a function of the distance between this package and their ideal points: the shorter this distance, the higher their utility. If proximity models can explain voters choices at the ballot box, then it is reasonable to assume that they can also make similar assessments when evaluating democratic performance. That is, a winner would be more satisfied with how democracy works because by voting for a winning party/candidate she can enjoy policy outputs that are comparatively closer to her preferences compared to the ideal point of a loser. We call this the policy effect of being a winner (PE). As a result, if PE effect matters, then we would expect that voters who backed losing parties will feel better if the eventual winner is closer to their own self-placement on the left-right scale than if the government is less proximate. In the former case, disutility in terms of policy implemented by the government is lower, which assuages the negative impact of being a loser on satisfaction with democracy. Exactly the opposite should happen with a winner. In cases of high voter polarization, losers could become so disaffected as to consider alternative regime types (Ginsberg and Weissberg 1978). In sum, greater distance between voters and the winning party exacerbates the winner-loser disparity. If they are close together, then this gap could be minimized. We therefore present our first hypothesis: Hypothesis 1: the policy effect (PE) of winning. Winners (losers) who are closer to the ideal point of the current government will be more satisfied with the way democracy works than winners (losers) who are located farther away. Clearly, the cabinet s position on the underlying ideological dimension of political competition matters a lot in this regard. Indeed, as long as the distribution of voters is approximately symmetric around some central value, as is usually the case (see Ezrow 2008; this is also true for our sample: see below), then an ideologically moderate cabinet, by minimizing the average distance between its position and the position of the whole electorate (irrespective of the voters winner/loser status), should contribute positively to the overall level of satisfaction with democracy, at least through an indirect effect on the relevance of being a winner. The opposite should happen when a cabinet takes a relatively radical ideological position. b. The Absolute and Marginal Dynamic Effect The winner-loser effect can also present a dynamic pattern given that in a democracy today s winners could be yesterday s losers, and vice versa. This dynamic can be viewed (and analyzed) from two alternative perspectives: an absolute pattern on one side, and a marginal one on the other. This difference relates related to how past (present) experience affects present (past) experience. If we assume that people prefer more to less (for a similar perspective see Anderson and LoTempio 2002), including in terms of attitudes toward the political system, then we can expect the following relationship: winning more should lead to higher levels of satisfaction with democracy than winning less; conversely, those who experience only losses should be the least satisfied, those who experience some loss and some gain across elections 4

should be in the middle, and those who experience only victories should be most satisfied. Moreover, among citizens in the middle category (mixed electoral results), we expect that those who won in the latest election (but lost in the previous one) express greater satisfaction with democracy than those who won in the past but lost in the most recent round. This makes sense if we assume that individuals discount more distant personal experience compared to more recent ones. Let WW stand for a double-winner (i.e. who voted both in the present and previous elections for a winning party/candidate), and LL a double-loser, with WL as someone who won in the present election but lost in the previous one, and LW a person who lost in the present election but won in the past. Our discussion leads to the expectation that in terms of satisfaction with democracy, we should find the following lexicographic ranking: Hypothesis 2a: the absolute dynamic impact (ADI) of winning. WW > WL > LW > LL To our knowledge, scholars have not examined this intuitive, albeit crucial, point in depth. Anderson and LoTempio (2002) is the only work that addresses a similar subject by trying to model the dynamic pattern involved in the winner-loser dichotomy. Specifically, they analyze if voting for the presidential and the congressional winners has a cumulative effect on levels of political trust of American voters. Using data from the 1972 and 1996 US presidential and congressional elections, they found that those who voted for the presidential winner were significantly more trusting than those who did not. In contrast, voting for the congressional majority party did not affect levels of political trust. Moreover, when they examined the differences in trust among double presidential and congressional winners, winners in only presidential or congressional contests, and double losers, they found that partial winners were more trusting than double losers, and double winners most trusting of all. In studying the impact of the winner-loser effect on satisfaction with democracy, it seems more natural to compare the status of being a winner in the present election with that of being a winner in the previous election. This implies that in a presidential system, we should compare the results across two consecutive presidential races, 2 while in parliamentary systems we compare results of two successive legislative elections. However, the dynamic pattern involved in the winning-losing effect can also be analyzed from a different perspective, which focuses more on its marginal rather than absolute consequences. That is, it can be reasonably assumed, following the standard theory of utility, that citizens derive a decreasing marginal utility in winning (and, for similar reasons, an increasing marginal utility in losing). In other words, what could really matter for voters is to have their views represented by the government at least once. If this happens, then the relationship between winning (losing) in the present and past is no longer an additive relationship (as previously assumed), but a conditional one instead. Consequently, the impact of winning today should matter more for those who have not enjoyed this experience in the past, and vice versa for losers who have never tasted defeat heretofore. In contrast, for voters who have previously been winners, winning today should matter far less in influencing their degree of satisfaction with democracy. We can therefore derive the following hypothesis regarding the dynamic nature of the winner-loser effect: 2 We also follow the same approach for semi-presidential systems. 5

Hypothesis 2b: the marginal dynamic impact (MDI) of winning. The marginal effect of winning in the present (past) election should matter more for citizens who lost in the past (present) election. Once again, which of the two dynamics is more appropriate in explaining the impact of being a winner has important normative and political implications. If the true dynamic impact is a marginal one, then a political framework that experiences frequent cabinet alternation represents the most favorable scenario to increase the diffusion of satisfaction with democracy, given that it will increase the proportion of the electorate who has experienced winning. On the other hand, if the absolute dynamic were the true one, then this would not necessarily be the case. The next section is devoted to detailing the information needed for an empirical test of our hypotheses. 3. Data and measurement issues To control for both hypothesis we use the second module (2001-2006) of the Comparative Studies of Electoral Systems (CSES) project. 3 This dataset employ exactly the same battery of questions across countries, and constitutes the largest single source of available cases. In particular, we include in our analysis all countries deemed free according to Freedom House at the time the election covered by the survey was held, as well as at the time of the previous election (this is necessary given that we are going to compare the winner-loser effect in both the current and the previous elections). This leaves us with 31 countries (see Table 1 below), a considerably larger pool compared to previous researches, covering not only Western Europe, but also Eastern Europe, South America and Asia. 4 Table 1 Another advantage of using CSES data is that questionnaire items include the necessary information needed to build our variables. In particular, we know: a) the parties (or presidential candidates) for which the respondent voted in the current and in the previous election; b) respondent self-placement on a 11 point left-right scale; 5 c) citizen placements of political parties (or presidential candidates) on the same scale. 6 Using the information coming from a) we can classify a citizen s winner/loser status in both the current and previous elections. A respondent is considered as a current winner if she voted for a party that belongs to the cabinet at the moment in which the CSES survey was administered (and the information regarding the level of satisfaction 3 Data and documentation can be downloaded from http://www.cses.org. 4 We did not include Philippines given the lack of information needed to estimate several of the independent variables controlled for in our analysis. 5 The Japanese survey used a Progressive Conservatives scale as an equivalent to Left Right. 6 In the case of Presidential and Semi-Presidential systems, whenever the left-right scores for the Presidential candidate winning the election was not presented in the surveys, we used the left-right score of the party to which that candidate belongs to. For Belgium, we use party positions from the Benoit and Laver expert-survey (Benoit and Laver 2006), given that this information is lacking in the CSES survey. 6

with democracy was recorded). Usually, it coincides with the first cabinet in power after the election. A past political winner is estimated in a similar way. In this case, however, we considered all the parties that had a chance to enter cabinet during the life of the previous legislature. 7 Columns 2 and 3 in Table 1 report years for the current and previous elections under analysis, while column 4 shows the type of election (legislative or presidential). Finally, column 5 reports if there has been any major government change (or intergovernmental change) following each of the two elections considered (see Yi Feng 1997). Quite interestingly, among the countries that did not experience any major governmental change in the current election, only 22 percent of respondents who were losers in the previous election emerged as winners in the current one. In contrast, in countries that did experience a major governmental change in the current election, this percentage increases threefold to 65%. This difference points to a well-known substantial stability in voting behavior. In this situation, therefore, crossing the winner/loser divide is usually due to the changing fortunes of a voters supported party than a change of her own partisan preference. 8 To test the impact of ideology, we construct a measure of spatial distance (labeled proximity ) between voters and government. The fact that estimates of party positions and voter spatial locations are calculated from the same data source eliminates problems of comparability between voter and party placements. 9 Specifically, our proximity variable (i.e., the spatial distance between elector i and cabinet j) is estimated in the following way: PROXIMITYij = xij Pj (1), 7 When we had a double-ballot to elect a president, we classify as winners citizens who reported voting for the winning candidate in the first round. Whenever the survey also contains information on the reported vote in the second round (this is true for Brazil, France, and Romania, but not for Chile and Peru), we replicate the analysis by coding as winners citizens who actually voted for the winning candidate in the second ballot. This does not yield significantly different results. Unfortunately, similar detail on first and second ballot vote in the previous election is not available. Finally, we only include in our analysis respondents who meet voting age requirements in each country at the time of the previous election as reported in Table 1. For example, for Italy we exclude from the analysis all respondents that at the time of the 2006 election were younger than 23 years old (meaning that in the previous election in 2001 they were not eligible to vote). We exclude these respondents since their winner/loser status at the previous election cannot be estimated. 8 As a check on the accuracy of the reported vote, we compared the aggregate distributions of actual and past election outcomes and recalled election outcomes. In both cases, we found a tendency of over-representation by respondents favoring the victorious party, though the difference is not large. In the current election, 49% of respondents reported voting for the government parties in their countries; this is slightly higher than the 45.9% who actually voted for these parties. The same is true for the previous election: 50.4% of reported votes versus 49% of actual votes. 9 We also estimate cabinets left-right position using two expert surveys by Benoit and Laver 2006 and Wiesehomeier and Benoit 2009. Our main findings are not affected by the source of party placement data used(a result that is not surprising, given the high correlation between mass and expert scores). 7

where: x ij = the ideal point of elector i in country j along the Left-Right spectrum; P j = the position of cabinet j s, estimated as the weighted average position of the cabinet parties. Considering a weighted average to compute the left-right position of the cabinet reflects the fact that declared government policy positions closely tracks the weighted mean position of cabinet parties (see Warwick 2001). 10 Table 2 reports the value of P j across countries in our sample, together with the average value of PROXIMITY under three different scenarios: considering all respondents, and differentiating them by winner/loser status in the current election. As expected, the value of PROXIMITY is considerably lower among winners than losers in most cases (the only exceptions are Canada, Peru and Taiwan, where the two values of PROXIMITY for winners and losers are similar). In particular, we find the widest difference in Italy (4.02) and Hungary (3.07). We also include in the Table a general measure of ideological eccentricity of each cabinet, defined as the absolute distance between P j and x j (i.e., the average ideological position of the respondents in country j). The correlation between this measure and the values of PROXIMITY reported is particularly high for losers (.80), but also not negligible for winners (.13). This implies that the more the left-right position of a cabinet diverges from the ideological centre of the electorate, the more the overall value of PROXIMITY increases. Indeed, in every country in our sample, the ideological distribution of voters is always approximately normal around a central value that lies in a relative short ideological range: from 3.4 in Poland to 6.1 in Ireland. 11 Table 2 Our dependent variable is the level of satisfaction with democracy (SWD). As in many previous studies, this concept was measured by the question: On the whole, are you very satisfied, fairly satisfied, not very satisfied, or not at all satisfied with the way democracy works in {country}? We reversed the original coding so that, on a 1 to 4 scale, higher values correspond to greater satisfaction with democracy. Since our dependent variable is a four-point scale, a linear regression assuming interval level measurement is not an appropriate method of analysis. An alternative strategy would be to collapse the four original categories comprising the SWD variable 10 We also estimate two different variants of PROXIMITY. First, we have estimate P k by simply averaging positions of all cabinet parties. Secondly, we utilize a quadratic utility function. Neither approach affects our main findings. 11 Chile represents an outlier, given that the value of x j in this case appears to be quite radical (i.e., an average value of less than 2). However, our results are not changed when the analysis is replicated after dropping the Chilean case. 8

into two, creating a binary variable. However, this option would fail to take advantage of potentially useful information. 12 The method most suited to our dependent variable is therefore an ordered logit model. In ordered logit, an underlying score (or latent variable) is estimated as a liner function of the independent variables and a set of cutpoints (or thresholds). The probability of observing a given outcome therefore corresponds to the probability that the estimated linear function is within the range of the cutpoints estimated for the outcome. In our case, there are four possible outcomes or categories (not al all satisfied, not very satisfied, fairly satisfied, very satisfied). Under this scenario, we are interested in identifying at which point of the latent scale the category not very satisfied changes to fairly satisfied (and similarly for the other categories). This is labeled the cutting point. Moreover, note that our dataset is hierarchically organized, with one level (individual respondents) embedded within another level (countries). Ignoring the multilevel character of data may generate possibly incorrect standard error and inflate Type I error rates. Furthermore, this violates the assumption that errors of the different observations are independent (Steenbergen and Jones 2002). Therefore we decide to estimate a multilevel model, allowing for each observation to be correlated within country, while being independent between countries. This allows us to judge the relative importance of individual as well as contextual covariates and avoid possible misinterpretation in terms of ecological or atomistic fallacy (Courgeau 2003, Bryk and Raudenbush 1992). The formulation of the multilevel ordered logistic regression we adopted is given by the following equation 13 : { ( x ζ1 )} β2( 0 ) β3 β4( 0 ) ( t WINNER ) ( t WINNER ) ( t WINNER ) X Z k logit Pr y > s, = t WINNER + PROXIMITY + t WINNER PROXIMITY ij ij j ij ij ij ij + β + β + β + β + ζ + ε - (2) 5 1 ij 6 0 ij 1 ij w ij z j 1j ij s where: Pr y > s x, ζ is the probability that respondent i living in country j presents a level ( ij ij 1 j ) of satisfaction with democracy higher than the threshold s; twinner is a dummy variable that assumes values of 1 if citizen i living in country j 0 ij voted for a winning party/candidate in the current election; 12 As a control check, we also estimate a model by creating a dichotomous variable. Results obtained are very similar to the ones presented below. These results are available upon request from the authors. 13 To our knowledge only in Halla et al. (2008) a multilevel ordered logit has been employed to analyze the determinants of the level of satisfaction with democracy. Wells and Krieckhaus (2006) as well as Anderson and Tverdova (2003) employ a multilevel model but they treat the dependent variable as it were an interval-level one. Note that the covariate effects in (2) are constant across categories, a property referred as the parallel regressions assumption. Using the method proposed in Rabe-Hesketh and Skrondal (2005) we found that this assumption appears to be reasonable in our case. 9

PROXIMITY ij is a variables that measures the distance between citizen i and the current government policy preference in country j according to (1); t 1WINNERij is a dummy variable that assumes values of 1 if citizen i living in country j voted for a winning party/candidate in the previous election; w ij are vectors of individual control variables (see below); z j are vectors of country control variables (see below); ζ is the overall intercept of the cumulative logits varying over countries j; 1 j ε ij is the error term unique for each i 14 ; k s are the thresholds of the ordered logit (3 in our cases: that is, the number of categories of our dependent variable minus 1). The two interaction terms presented in (2) are crucial to testing our hypotheses. In particular, according to Hypothesis 1, in order for the PE hypothesis to matter in addition to the partisan effect of winner status, the marginal effect of being a winner in the present, that is β 1 + β 3 * PROXIMITYij, should increase (decrease) as the value of PROXIMITY ij increases (decreases). In other words, we expect a negative and significant β 3 coefficient. The second interaction term presented in (2), on the other hand, allows us to explore two possible alternative dynamic effects of being a winner. A significant and negative β 5 coefficient would alert us to the existence of a marginal dynamic impact, while any absolute dynamic impact should be highlighted by a positive and significant coefficient for β 1 and β 4 but not for β 5. Figure 1 summarizes our data by plotting the average percentage of respondents who express satisfaction with democracy in each country under analysis (ranked by the magnitude of this percentage). On average, 64 percent of respondents indicate satisfaction. However, the figure shows considerable cross-national variation in attitudes toward democracy: positive evaluations range from 28 percent in Bulgaria all the way the 95 percent in Denmark. 4. Control variables Figure 1 We draw on the large empirical literature on the determinants of satisfaction with democracy and control a number of variables at both individual and country levels that have been found significant in cross-country analyses (e.g. Canache, Mondak, Seligson 2001; Mattes and Bratton 2007). (See the Appendix for a list of control variables). With respect to the former, in addition to socio-demographic factors (e.g. gender, age, education), we include variables measuring the existence of respondent s party identification (a dummy variable coded 1 if the respondent considered herself close to a party and zero otherwise) and evaluations of government overall performance to ensure that the winner-loser variable is not simply a proxy for the effect of partisanship or a general identification with those in power. We also controlled for three 14 Therefore ξij = ζ1 j + εij is the total residual. 10

variables related to the political behaviors of the respondents. The first two variables concerns rather orthodox forms of political participation, while the third one refers to more heterodox forms. Namely, we include a dummy that captures if the respondent voted in the current election, a dummy that records if the respondent attempted to persuade others to vote for a particular party or candidate, and finally, a dummy that takes values of 1 if the respondent has taken part in protests or demonstrations. 15 The last two variables that we introduced at the individual level relate to respondents general political attitudes: the first captures how much respondents believe that the people in power could make a difference, while the latter takes a value of 1 if respondents agree with the statement that democracy is better than any other form of government. Regarding the variables related to the aggregate (i.e., country) level, we controlled for: a) the quality of formal institutions. The idea here is that institutions that promote the quality of resource allocation and provision of public good increase SWD. We take institutions to broadly mean rules of the game in a society (see Wagner, Schneider and Halla 2008) and we use as their proxy a variable related to the quality of the enforcement of the rule of law in each country (see Kaufmann et al. 2002). b) the economic performance of a country. The worse (better) the overall economic performance of a country, the lower (higher) satisfaction with democracy is expected to be. We include in this regard two variables: first, the average GDP growth in the last five years; second, the deviation of GDP growth in the year immediately preceding the survey from the average GDP growth. This allows us to record the impact of last year s GDP growth above and beyond its contribution to the country s average GDP growth. By including both variables, we can differentiate between the short and the medium-term impact of the economic performance on SWD. c) a dummy variable that contrasts new democracies from long-established ones. We consider a democracy as established if it did not experience any interruption in democratic status in the last 25 years. Spain and Portugal, which became democratic only in the 1970s, are therefore regarded as old democracies; d) since our sample includes both parliamentary and presidential systems, we include a dummy that assumes a value of 1 for the latter form of government. This is a political institutional variable that has been rarely (if at all) controlled in the literature. e) finally, given that some works found a relationship between the level of satisfaction with democracy and the existence of an institutional framework that favors a consensual style of decision-making vs. a majoritarian style we included a variable that capture the strength of the institutional checks and balance (and therefore the diffusion of power: see Keefer 2007) in each country as well as the Gallagher Index of disproportionality of the electoral rules (Aarts and Thomassen 2008, Anderson 1998). 16 We can consider both variables as proxies of the difference between consensual and majoritarian democratic systems. Indeed Lijphart (1999) makes the point that consensual democracies outperform majoritarian democracies in terms of responsiveness and do at least as well in terms of efficiency. This should lead to higher 15 On the difference between orthodox and unconventional forms of political participation, see Barnes and Kaase 1979, Jennings et al. 1990. 16 For Presidential and Semi-Presidential system, we estimate the Gallagher Index following the method proposed by Lijphart (1999), i.e., by taking the geometric mean of the Gallagher Index computed for presidential and for legislative elections. 11

levels of satisfaction with democracy (Wagner, Dufour and Schneider 2003). Similarly, Miller and Listhaug (1990) argued that a smaller number of parties (primarily deriving from the non-proportionality of electoral rules) is correlated with lower satisfaction with democracy because there are fewer policy choices for citizens. In other words, because some electoral systems inhibit the emergence of new parties to take account of new demands, system support may decline in the long run. Thus, we expect lower levels of satisfaction in countries with a high value of the Gallagher Index. Moreover, we also control for the existence of any mediating impact of consensual (majoritarian) systems on the relationship between winner/loser status and satisfaction with democracy. It has been noted that the nature of representative democratic institutions in which losing and winning takes place is also important. For example, Anderson and Guillory 1997 (see also Banducci and Karp 2003, Wells and Krieckhaus 2006) showed that those supporting opposition parties were likely to be less dissatisfied in consensual systems than their counterparts in presidential systems. The intuition is that even losers in consensual systems can influence policy to some extent, giving winners less absolute control than in majoritarian systems. In this regard, we introduce two interaction terms in the model, between winner status in the present election and the two aforementioned proxies for consensual democracies. 5. The analysis Table 3 reports the results of the three models we estimated. The first, null model is a model with no level-1 or level-2 predictors. The utility of this a-theoretical model is that it allows to decompose the total variance in our dependent variable between individual and country levels. Through this we can estimate the intra-class correlation, ρ, that is the variance in the level of satisfaction with democracy that can be explained by ρ = Var ζ Var ζ + ε. information on the country-level. Formally: ( 1j) ( 1j ij) Knowing that the variance component at the individual level (i.e., Var ( ε ij ) ordered multilevel logit is normalized to π 2 3, and that the value of Var ( ζ 1 j ) ) in a, shown in the fifth to last row in the table, is.67, we can infer that country-level variance in our data constitutes around the 18% of the total variance. Given that the data are measured at the individual level, the fact that country-level variance is proportionally much smaller than individual-level variance is not surprising (Steenbergen and Jones 2002). Moreover, the variance at level-2 is significant at 99% level. These results strongly lend support to the utility of applying multilevel analysis on our data (and to the possible perils of neglecting their hierarchical structure). Table 3 Next, we turn to the question of whether the models we have specified can account for the variance in SWD, beginning with Model 2. Regarding the control variables at both individual and aggregate levels, the results are largely expected (with one exception noted below). At the individual level, the respondents who are male, highly educated, possess party identification, and express a positive attitude toward democracy (not only agreeing that democracy is better than any other form of government, but also believing that who is in power can affect how a 12

country is ruled) have a higher probability of being satisfied with democracy. Interestingly, both forms of orthodox political participation are insignificant. In contrast, unconventional forms of political participation, as well as a negative assessment on the cabinet s performance, are significantly and negatively related to SWD. The age of respondents do not exert any significant impact. At the aggregate level, the quality of formal institutions (the RULE OF LAW variable) is positively related to SWD. The same holds true for economic performance, albeit short-term economic performance (GDP) seems to influence democratic satisfaction more than medium-term economic performance (AVG GDP). New democracies display a lower level of SWD compared to long-established democracies, confirmed that losers have not yet learned to lose in countries where democratic governance is of such recent vintage. (Anderson et al. 2005: 108). On the other hand, electing the head of the executive directly is positively related to SWD (albeit only at the 90% confidence interval). Finally, the checks and balance variable (CHECKS) is highly significant and positively related to SWD as well, while the Gallagher Index is clearly not statistically distinguishable from zero. 17 We can now focus on our main hypotheses. For this purpose, the two interactive terms involving the twinner 0 variable play a crucial role. Let us start with the hypothesis related to the dynamic impacts of being a winner. The highly significant (and negative) coefficient for the interaction term between twinnerand 0 t 1WINNERij clearly indicates the existence of a marginal (rather than an absolute) dynamic impact of being a winner. Figure 2 in this regard plots the marginal effect of winning in the past as a function of twinneron 0 the probability of being fairly or very satisfied with democracy (that is, Pr ( yij > 2 x ij, ζ1 j ) ), holding other independent variables (including PROXIMITY) at their mean level. As the figure shows, winning in the previous election significantly and positively affects the probability of being satisfied with democracy only for losers in the present election. By contrast, there is no cumulative effect in the winning experience: being a winner in the past is not significant if the respondent has also won in the current election. Figure 2 Note that in Model 3 there is a second interaction term involving twinner. 0 Therefore, to assess the true marginal effect of winning in the present (and its temporal dynamic: see Brambor et al. 2006) we also need to control for the value of PROXIMITY, that is for the policy effect of winning. This is what we have done in 17 By introducing the aggregate-level variables, we explain a substantial proportion of the country-level variance (given that it reduces from.67 to.12: -82%). We can therefore claim that we have minimized a potential omitted variable bias (see Anderson and Tverdova 2003). However, both in Model 2 and 3, the estimated intra-class correlations, ρ, implies that around 4% of the total variance in the level of satisfaction with democracy is still explained by information on the country-level. To judge whether this data structure necessitate a multilevel estimation, we can calculate the design effect, defined as 1+(average cluster size-1)* ρ. A design effect greater than 2 indicates that the clustering in the data needs to be taken into account (Muthen and Satorra, 1995). Our results is 25.6. Ignoring the multi-level nature of the dataset, even after having introduced the aggregate-level variables, simply is not advisable. 13

Figure 3, which depicts the marginal effect of being a winner in the present on the probability of expressing satisfaction with democracy across the observed range of PROXIMITY in our dataset. Moreover, in Figure 3(a) we fix the value t 1 WINNER ij at zero, while in Figure 3(b) we consider respondents who won in the past election (that is for t 1WINNERij equals to 1). We also superimpose a histogram portraying the frequency distribution for PROXIMITY over the marginal effect plots (the scale for the distribution is given by the vertical axis on the right-hand side of the graph). The results are very intriguing: in the former situation, the marginal effect of twinneron 0 the probability of being fairly or very satisfied with democracy increases as the value of PROXIMITY declines. This marginal effect ceases to be significant for values of PROMIXITY higher than 6. This is however a condition satisfied by almost all respondents in our sample (95.6%). On the other hand, for respondents who won in the past, being a winner in the present matters, but only when PROXIMITY is less than 2, a condition much more stringent than the previous one. Indeed, this condition is satisfied by just the 53% of our sample. 18 Moreover, note also that for values of PROXIMITY lower than 2, the marginal effect of being a recent winner is clearly higher in Figure 4(a) than in Figure 4(b). For example, the marginal impact of being a winner in the present when PROXIMITY equals 1 increases the probability of being satisfied with democracy by 9.6% (95% c.i.: 7.5% -11.8%) in Figure 4(a) and by three times less in Figure 4(b) (+3.3%; 95% c.i.: 1.4%-5%). This clearly confirms the marginal (rather than absolute) dynamic effect that we found in Figure 2 concerning t WINNER 1 ij. Figure 3 Thus, our hypothesis that greater (lesser) distance between voters and the government can exacerbate (alleviate) the winner-loser disparity finds empirical support in the analysis, especially when we consider respondents who have not won in the past. In other words, a spatial reasoning lies behind the relationship between winner status (in the present) and SWD, and this aspect allows us to better understand the dynamic impacts of being a winner. Moreover, and quite interestingly, comparing the marginal effect on SWD of winning in the present (for respondents who lost in the past) and of winning in the past (for respondents who lost in the present) for similar values of PROXIMITY yields similar results. In the former case, the probability of being fairly or very satisfied increases by 8.1% (95% c.i.: 6.2%-10%); in the latter case, +5.7% (95% c.i.: 3.7%- 7.8%). That is, voters do no seem to dramatically discount the impact of being a winner on satisfaction with democracy (provided the respondent has won just once)., Finally, Figure 4 plots the marginal effect of PROXIMITY depending on twinner 0 (holding all other variables at their means) to fully test conditional relationship suggested by Berry et al. 2009. The figure demonstrates that decreasing the PROMIXITY value from the median to its first quartile does not affect the probability 18 We also note that for some extreme value of PROXIMITY, i.e., larger than 9, the marginal effect of twinnerbecomes 0 negative. However, we have no observations that satisfy this requirement. We therefore disregard this result as irrelevant. 14