Adrift or adroit? A panel study in 58 waves on the sources of electoral volatility in the Netherlands,

Transcription

1 Adrift or adroit? A panel study in 58 waves on the sources of electoral volatility in the Netherlands, Tom van der Meer - University of Amsterdam Rozemarijn Lubbe - University of Amsterdam Erika van Elsas - University of Amsterdam Wouter van der Brug - University of Amsterdam Paper prepared for the panel The Electoral Cycle Revisited ECPR General Conference 2011 Reykjavik, August 24 August 27, 2011 First draft Please do not cite without explicit authorization by the authors All comments welcome Contact: Tom van der Meer t.w.g.vandermeer@uva.nl

2 1. Introduction Electoral volatility is increasing in most western democracies (Bartolini & Mair 1990; Dalton et al. 2000; Drummond 2006; Mair 2008). Subsequent election results show that electoral support for political parties is no longer a given. The number of voters shifting from one party to another has risen. The Netherlands exemplify this trend: once a typical example of a pillarized society (Lijphart 1968) with frozen cleavages (Lipset & Rokkan 1967), the country is now a typical example of a volatile democracy (Mair 2008): four of the last five Dutch parliamentary elections (of the Lower House) belong to the most volatile elections in Western Europe since But what does the increase in volatility mean? On the one hand, electoral volatility is often treated as an ominous phenomenon. It would lead to ineffective government, populism and unstable democracy as politicians continuously need to adapt to the whims of their constituency. This description of a volatile electorate would hold if voters could be characterized as increasingly whimsical (Walgrave et al. 2010), fashionable (Andeweg 1982) or outright as drift-sand (as various politicians commented) that do not decide in an informed way but through passions of the heart or possibly even worse due to short-term media images. On the other hand, electoral volatility also has a positive connotation, especially in comparison to the outright dystopian image of a society without any changes in vote intentions. Exactly because voters support is not a given, politicians and their parties remain accountable to the electorate. The positive perspective on volatility thus implies an emancipated electorate consisting of sophisticated voters (Dalton 1984): more and more voters have the necessary political skills to make political choices without relying on traditional loyalties (Lachat 2005). At the very minimum, some extent of electoral volatility is thus required for a stable democratic system. Although electoral volatility is at the very heart of political science research, it has remained unclear which of these two perspectives captures the nature of electoral volatility. To understand the nature of electoral volatility, we need to understand which voters are most likely to be volatile. Does electoral volatility reflect uninformed and capricious voters disinterest and lack of commitment (i.e., are voters adrift) or does it reflect emancipated and critical voters skilful adaptation to changing circumstances (i.e., are voters adroit)? One reason of this inclarity is the continued lack of individual panel data to study electoral volatility (Bartolini & Mair 1990; Roberts & Wibbels 1999). Most studies on volatility are aggregate studies that focus on net volatility, that is: the changes in vote shares of political parties during elections. Crossnational and longitudinal studies on net volatility have provided structural and institutional explanations such as electoral system and economic development rates (cf. Bartolini & Mair 1990; Roberts & Wibbels 1999; Tavitz 2005). Yet, such analyses cannot be used to study the individual sources of volatility such as education, income, and associational involvement; any attempt to do so using aggregate figures risks an ecological fallacy. Moreover, aggregate shifts in parties vote shares (net volatility) is by definition an underestimation of all the individual changes that take place within the electorate (gross volatility): opposite changes cancel each other out in aggregate studies. Individual level studies of electoral volatility are scarce (Kuhn 2009). Moreover, although highly informative, they were limited by various mostly methodological problems. First, previous individual level studies often had very few points of measurement, they were limited to a comparison of actual votes at subsequent elections, or at best vote intentions during the month(s) before an election. Hence, voters who make a single (informed) change in preference could hardly be pulled apart from voters who change preference more often. Second, several studies had to assess past behaviour through retrospective questions, despite all the substantial biases towards consistency (Beasley & Joslyn 2001; Van Elsas et al. 2011). To further the debate on electoral volatility, we require a panel study with more waves that span the full electoral cycle including the years in which there were no parliamentary elections. Such panel data allow a distinction between single, informed changes and capriciousness. In this study we therefore employ an extensive panel data set that allows us to open the empirical black box, the Dutch 1Vandaag Opinion Panel (1VOP). These panel data cover adult respondents who participated in at least 2 of the 58 panel waves between November 2006 and June This enables us to analyze the individual level sources of electoral volatility. We use these panel data to answer the following research questions: 1. To what extent are Dutch citizens volatile in their vote intentions? 2. To what extent is electoral volatility explained by (a) voters years of socialization, (b) voters level of (political) sophistication, (c) voters embeddedness in family and associational life, (d) voters socio-economic vulnerability, and (e) voters use of media?

3 3. To what extent do these explanations differ with the timing and the direction of electoral volatility? This study aims to fill the theoretical and empirical gap in our knowledge on electoral volatility. By answering these three research questions, we aim to shed light on its nature and individual level sources. 2. Theory and hypotheses 2.1. Volatility in the Netherlands The Netherlands are an intriguing case to study gross electoral volatility for several reasons. In the multiparty system voters have the highest potential to change vote intention, because both the high number of ideological spaces between parties (Lachat 2007) and the relatively small size of these ideological spaces (Roberts & Wibbels 1999) stimulate volatility. Moreover, competitive systems (such as the Dutch party system since 2002) also stimulate volatility, as political elites tend to outbid each other (Sartori 1976). Consequently, the Netherlands have become the established democracy with the highest level of net volatility in the last decade(s). The elections for the Dutch Lower House (Tweede Kamer) have been the most volatile of Western Europe since Figure 1 describes the net volatility as the share of votes that all winning parties gained compared to the previous election (Mair 2008). Note, that in the highly proportional electoral system of the Netherlands the shifts in votes are highly similar to the shifts in parliamentary seats (Anker 1994). Figure 1. Net volatility 1 : aggregate changes between elections 1 Net-volatility is the sum of the aggregated electoral growth of all winning parties compared to the previous election (Pedersen 1979) Source: Mair 2008; own calculations. Until the late 1960 Dutch elections were highly predictable more so than the Western European average. In the pillarized society voters party preference could to a large extent be predicted by their socioeconomic and religious backgrounds (Irwin & Van Holsteyn 1989). Because of the close ties between political leaders, parties, and their pillarized electorates, true electoral competition between political parties remained very small: on average the election outcomes changed merely by 5 per cent. Demographic changes played a large part in these changes. Electoral volatility increased with depillarization in the late 1960s. Between 1967 and 1986 the average electoral changes was about twice as high as in the previous decades. Dutch elections were considered to show large variations, while Dutch voters were characterized as a fashionable electorate, susceptible to changing public moods (Andeweg 1982). The increase in volatility fits the general western pattern since 1970 (Drummond 2006). Yet, the following decades would show that the increased volatility of the 1970s and 1980s would only be a precursor to even larger changes in the 1990s and 2000s. When net volatility increased to 20% in the 1994 elections, political commentators spoke of a political earthquake and an unpredictable electorate. When the elections of 2002, 9 days after populist party leader Pim Fortuyn was murdered, showed a level of volatility above 30%, commentators spoke of a political revolt. Since that peak volatility declined a bit, but remained at a historically and comparatively very high level Net and gross volatility: 8 measures of volatility The net volatility in Figure 1 only provides macro-level information. It is not informative about the individual level processes that underlie these macro changes. Hence, our understanding of the meaning of

4 electoral volatility remains rather obscure. Voters of losing parties may not have gone directly to winning parties. 1 Small net changes might even be the result of large gross changes, when opposite streams of voters cancel each other out. Net volatility rates are therefore always a not necessarily systematic underestimation of gross volatility. Previous studies on gross electoral volatility focused almost exclusively on changes in vote intentions that took place between subsequent elections or during the campaign months before an election (e.g. Granberg & Holemberg 1991; Blais 2004; Lachat 2007; Söderlund 2008; Walgrave et al. 2010; but also see Kuhn 2009 for an exception). Yet, voters preferences do not solely change during those months. Rather, electoral volatility may even be larger between elections than during election time. In the Netherlands, for instance, the new party TON gained and lost nearly 30 virtual parliamentary seats within a few months in 2007 and Similarly, in opinion polls PvdA held 60 virtual seats in the Spring of 2006, went down to 15 seats in December 2009, and recovered some in February and May Hence, a thorough understanding of electoral volatility should encompass a full electoral cycle. A comparison with the literature would benefit from a comparison of election volatility (changes in actual vote choice during elections) with campaign volatility (changes during preceding campaign periods) and general volatility (outside campaigns and elections). Moreover, we need to distinguish between voters who display single, substantive changes in vote intentions on the one hand and those who change vote intentions repeatedly. If we define volatility as behaviour that displays a likelihood to change vote intentions more often, it matters whether a voter decides to switch parties once or twice (and sticks to these changes afterwards) or whether a voter switches parties repeatedly. Previous studies on gross volatility were limited by their data: because most encompass merely two repeated waves (some through panel data, others through retrospective questions 2 ), they cannot distinguish between these types of volatility (e.g. Granberg & Holemberg 1991; Blais 2004; Lachat 2007; Söderlund 2008; Walgrave et al. 2010). Finally, when citizens vote intentions change, they may shift between ideologically similar parties (within a single block of parties) or between ideologically dissimilar parties. Inter-block volatility may come closest to the description of volatile voters as whimsical and drift sand: these changes appear unstructured. By contrast, intra-block volatility signals at least some extent of political sophistication, as voters shift between parties but stick to the wider party block. Generally, electoral volatility is characterized by intra-block changes rather than inter-block changes (Bartolini & Mair 1990; Granberg & Holmberg 1990; Lachat 2007; Walgrave et al. 2010; Van der Meer et al. under review). However, previous studies were restricted to the use of a single overarching left-right dimension to delineate party blocks. We build on a more nuanced, two-dimensional spatial model to distinguish between various party blocks (Van der Meer et al. under review). For these reasons, we distinguish between various forms of electoral volatility in this study. First, we assess three basic measures of electoral volatility: (1) whether or not voters changed their vote intention at least once during the electoral cycle; (2) the frequency with which they changed their vote intention; and (3) the number of political parties which they intended to vote for at least once during the electoral cycle. Second, we distinguish between (4) changes in actual vote choice during parliamentary elections, (5) changes in vote intentions during campaign periods before those elections, and (6) changes in vote intentions outside the campaign and election periods. Third, we distinguish groups of volatile voters by the degree they shift vote intentions (7) between ideologically similar parties (intra-block volatility) or (8) between ideologically dissimilar parties (inter-block volatility). This paper thus focuses on the questions which citizens are most likely to change their vote intentions across time. We analyse the presence and extent of these changes, as well as the timing and the directions of change. 2.3 Explanations of electoral volatility We propose various theoretical models to explain why some groups of voters are more volatile than others. The first model is the socialization model. With age, voters should become more rigid as they have been socialized in the party system (Campbell et al. 1960; Dalton 2000), becoming less responsive to new 1 A common misunderstanding of the Dutch electoral shift from 2002 to 2003 was that the disappointed voters of the List Pim Fortuyn (which lost 18 seats) went to the Labour Party PvdA (which gained 19 seats). Electoral studies of 2 The use of retrospective questions may be problematic as they suffer from random and systematic errors (Beasley & Joslyn 2001; Van der Eijk 2002; Van Elsas et al. 2011). Voters and especially volatile voters - tend to use their current vote intention as a cue to recall their past vote an effect that grows stronger with time since that past behaviour (Van Elsas et al. 2011).

5 cues (Walgrave et al. 2010). Indeed, various authors found that from election to election, younger voters are more likely to change their party preference than older voters (e.g., Schmitt-Beck et al. 2006; Söderlund 2008; Kuhn 2009). We come to our first hypothesis: H1. Young voters are more likely to be volatile than old voters. The second model is the cognitive sophistication model (Lachat 2007; Albright 2009; also see Kuhn 2009). It builds on the idea that citizens increasingly possess the level of political skills and resources necessary to become self-sufficient in politics (Dalton 1988). Cognitive sophistication entails political interest, knowledge and skills that voters can use to become politically active. Yet, the relationship between cognitive sophistication and volatility is not linear (Albright 2009). 3 Lachat (2007) argues that a crucial precondition of changes in vote intentions are the probability to receive political cues and the probability to yield to them (cf. McGuire 1968; Zaller 1992). Although they might be likely to adapt their non-attitudes to new information (Lazarsfeld et al. 1948), politically unsophisticated voters are less likely to receive and understand political cues, and therefore less likely to change their opinion. Politically highly sophisticated voters, by contrast, are likely to receive such political cues, but are less likely to accept these new messages if they are conflicting with their former preferences. Hence, in general, the most volatile voters should not be the unsophisticated nor the highly sophisticated voters, but the middle group (Lachat 2007; Kuhn 2009). The supposed balance between receiving cues and yielding to them invokes more specific hypotheses regarding the direction of changes. Moderately and highly sophisticated citizens are more likely to change between ideological similar parties, that is intra-block (Lachat 2007; Walgrave et al. 2010), as they are more likely to receive and understand information about the ideological structure of the party system. By contrast, lacking such information, less sophisticated citizens more likely to change at random; hence they are more likely to change between ideologically highly dissimilar parties, or inter-block. The second set of hypotheses reads: H2a. There is a non-linear relationship between voters level of cognitive development and volatility: volatility first increases, then decreases with rising levels of cognitive development. H2b. There is a non-linear relationship between voters level of cognitive development and intra-block volatility: volatility first increases, then decreases with rising levels of cognitive development. H2c. There is a negative relationship between voters level of cognitive development and inter-block volatility. The third explanatory model is the social embeddedness model. This model emphasizes that voters are politically socialized in their family environment (ref): stable family life would strengthen political attitudes. Similarly, voters choose to join associations with people of like-minded interests and attitudes. A process of selection and subsequent socialization also strengthens pre-existing attitudes (Hooghe 2007). 4 These effects are most notable in pillarized societies, where family life, associational life and church were so strongly integrated in social pillars, that electoral volatility was hardly present (Andeweg 1982). Classic sociological theories, building on Merton (1938), lead to the same hypothesis, albeit through somewhat different mechanisms. Social isolation (that is, absence of family ties or associational life) may cause feelings of anomie and alienation (a mismatch between aims and means and uncertainty about social norms). This in turn would stimulate electoral volatility. Our third hypothesis is therefore: H3. Voters who are socially more isolated (i.e., who have fewer social ties with family members, or within voluntary associations) are more likely to be volatile. The fourth model emphasizes socio-economic vulnerability, i.e., the socio-economic position of voters. On the one hand, we expect poor and unemployed to be more volatile than the rich and (self-)employed (Zelle 1995). Weak socio-economic conditions stimulate political distrust (Van der Meer & Dekker 2011), while 3 The classic cognitive mobilisation model did argue that the most sophisticated voters would be the most emancipated, the least aligned to political parties and therefore the most volatile (Dalton 1984). 4 The relationship between social embeddedness and volatility is more complex: the effect of social ties might be conditional on the congruence of political attitudes (Kuhn 2009). If there are different opinions in a household or in an association, voters are more likely to be volatile. Only a strong overlap in political attitudes might only reinforce these attitudes in a process of selection and subsequent socialization. Unfortunately, our data do not allow for a test of this effect.

6 this frustration (dissatisfaction with the current situation and cynicism about the future) is in turn a strong impetus to change ones vote intention (Adriaansen 2011, but also see Söderlund 2008 who finds no effect of disaffection). Yet, concurrently, modernization theory argues that not the lower class but the middle class is most likely to be volatile (Zelle 1995). The middle classes lack evident class-based political identities. In an environment where class-based voting is decreasing due to an increasing middle class, it would be exactly the lower class (i.e., the poor and the unemployed) and the upper class that will contain the fewest floating voters, and therefore be the least volatile. We formulate rivalling hypotheses: H4a. H4b. Poor and unemployed voters are more likely to be volatile than rich and employed voters. There is a non-linear relationship between voters level of income and volatility: volatility first increases, then decreases with rising income levels. Our final model revolves around media use. Media are a common scapegoat for the ills of society. They are indeed a prime candidate to stimulate electoral volatility (Latimer 1987). Supposedly, media generally prime voters attention primarily to short-term concerns, and thereby undermine the stability of the political agenda. Moreover, political news would strongly focus on persons, electoral horse races, and scandals (see Ansolabehere et al. 1991), thereby inducing societal discomfort (Elchardus & Smits 2002), political cynicism (Hendriks 2009, but see Adriaansen et al. 2011) and volatility (Kleinnijenhuis & De Ridder 1998). Low educated, political disinterested, and socially threatened citizens should be most susceptible to (changing) media images, and thus most likely to experience strong media effects on their level of volatility (Elchardus & Smits 2002). A more refined approach argues that not media use per se, but rather the type of media and the content in those media affect electoral volatility (De Vreese & Semetko 2004). Some media stimulate political knowledge more than others, which in turn might limit volatility in line with the cognitive sophistication model (cf. De Vreese & Boomgaarden 2006). Volatility may therefore be lower amongst people who are exposed to highbrow news media rather than predominantly entertainment media such as tabloids. Hence, our final set of hypotheses read as follows: H5a. H5b. Voters who read newspapers are more likely to be volatile than voters who do not. Voters who read tabloids are more likely to be volatile than voters who read highbrow newspapers. 3. Data 3.1. The 1Vandaag Opinion Panel We base our empirical analysis on the data from the extensive 1Vandaag Opinion Panel (1VOP), collected by the national television programme 1Vandaag. The data set covers 58 interview waves conducted between October 23 rd 2006 (a month before the 2006 parliamentary elections) to June 21 st 2010 (twelve days after the 2010 parliamentary elections). Each wave posed a question on the intended voting behaviour. As such, the 1VOP covers panel data on vote intentions over a whole electoral cycle. We restricted the sample to respondents participating in at least two waves (to allow for changes) and to respondents eligible to cast a vote, respondents on total. The frequency with which these respondents participated is quite high: half of the respondents participated in at least 15 waves Self-selection bias The 1VOP is put together on the basis of self-selection: respondents could sign up for the opinion panel and were subsequently invited for each web survey by 7. Neither the panel nor the samples per interview wave are thus randomly selected. Consequently, aggregate party preferences are not representative of the Dutch population. Yet, 1VOP is the best data set available for the purposes of our this article. 5 Since the 1VOP data set only reports age groups, we left out people in the lowest age group (younger than 20 when they first participated in the 1VOP). All analyses are restricted to those in the second lowest age group (20-27) and higher. This did not affect our findings substantively. 6 The average participation is 18.6 times;9% of the respondents take part in 40 waves or more. 7 Online accounts cannot be proven to be strictly personal. We used background questions on age, sex and educational level (that respondents were asked to fill in twice, in 2007 and in 2009) to check for potentially shared accounts: respondents with changing sex or decreasing age or educational level were excluded from the sample. These respondents made up less than 0.5 per cent of our data set.

7 Moreover, we should not exaggerate the self-selection problem. First, for the purposes of this study the self-selection bias would be hardly problematic. Although respondents may be more volatile than the Dutch population as a whole (due to priming effects of being in a panel) or less volatile (due to the stabilizing effect of repeated questions), there are no a priori reasons to expect that the explanations of electoral volatility might be mis-estimated in the 1VOP data set. Second, comparison of the 1VOP data to actual election outcomes show very similar results. The net volatility rate of the 2010 parliamentary elections was 23.7%, whereas net volatility according to the reported voting behaviour in these elections by the 1VOP-respondents differed only slightly (22.2%). If we match the election outcomes of 2010 to those reported in 1VOP, there is hardly a difference once we weight respondents according to their votes at the 2006 elections: both the outcomes and the changes in election outcomes correlate with 0.98 with those in the 1VOP data set. 8 Third, over- and underrepresentation in the sample would be problematic for multivariate analyses if certain types of respondents would be (nearly) absent. Thanks to an exceptionally large sample size, however, we can safely assume to cover a high variety of voters. A bias remains in favour of citizens who are interested in daily news and societal topics and who are willing to express their opinion on such topics. Nevertheless, 1VOP often deals with topics that are not particularly about (party) politics itself 9 ; hence there seems little overrepresentation of the politically interested Dependent variable: volatility General measures of volatility To assess electoral volatility, we use measures of respondents (self-reported) voting behaviour during the elections of 2006 and 2010 and their vote intentions during the 58 waves before, between, and after these elections. Vote intention was measured with the question: Which party would you vote for if parliamentary elections were held today?. The answer categories to this question cover all relevant parties at that time 10 (next to non-substantial categories to catch residual parties, non-voting, blank voting, and the common missing value categories). On the basis of the repeated item on intended voting behaviour we created variables that summarise electoral volatility over the whole period. First, we determined for each wave whether or not respondents substantially changed their vote intention compared to the last wave they participated in. Only shifts between political parties are counted as volatility; shifts to and from non-substantive answer categories (such as intending to vote blank, refusing to answer the question, or reporting to be undecided on the question) do not contribute to volatility. Hence, changing from party A (wave 1) to party B (wave 2) counts as volatility; changing from party A (wave 1) to being undecided (wave 2) does not. Changing from party A (wave 1) via undecided (wave 2) to party B (wave 3) leads to a volatility score of 1 change (at wave 3, when the substantive shift was made); changing from party A (wave 1) via undecided (wave 2) back to party A (wave 3) leads to a volatility score of 0 changes, as no substantive shift has been made. Based on these volatility scores per wave, we were able to assess the frequency of changes each respondent made. Similarly, we created dichotomous variables for each party to assess whether respondents intended voting for this party during the period under research. For each political party respondents score either a yes (intended at least once) or a no (never intended). Based on that information, we created a measure of volatility to assess the number of parties a respondent mentioned in total. Temporal measures of volatility Previous studies assessed volatility of the actual vote during elections, and/or volatility during the campaign period (Granberg & Holemberg 1991; Blais 2004; Lachat 2007; Söderlund 2008; Walgrave et al. 2010). Yet, opinion polls show that volatility outside the campaign and election periods is at least as large. We distinguish between three measures of volatility to assess whether the temporal delineation accounts for any differences with previous study. 8 On a related note, trends in vote intention in the 1VOP sample mirror those found in other polls during this period (Peil.nl, TNS/Nipo and Synovate). 9 Most waves of the 1VOP in this time span did not even include a question on vote intention; only the 58 in this study did, often as the final question in the questionnaire. 10 Christian Democrats (CDA), Labour (PvdA), the Socialist Party (SP), the Liberal-Conservatives (VVD), the Freedom Party (PVV), GreenLeft (GL), ChristianUnion (CU), the Democrats (D66), the Animal Party (PvdD), the Orthodox Christians (SGP), and the short-lived party Proud of the Netherlands (TON). TON never got elected into parliament, but in 2007 and 2008 had huge support according to the opinion polls. TON has been included as an answer category since the announcement of its foundation in November The LPF was only included in the four surveys before the 2006 elections. Since the LPF has won no seats in those elections, the party has not been included afterwards.

8 Election volatility, that is the change in the actual vote during elections, is measured by comparing the reported votes for the 2006 and 2010 elections of the Dutch Lower House (Tweede Kamer). Campaign volatility, that is the change during the campaign period, is measured dichotomously: do people change preference at least once in the time period directly after the fall of the Dutch cabinet in February 2010 (when it became clear that there would be early elections) up to the elections themselves in June Non-campaign volatility is also a dichotomous measure and concerns all changes in the timespan outside the Lower House election campaign, i.e. the changes from February 2007 up to early February Directional measures of volatility By far the most changes in vote intentions in the Netherlands between 2006 and 2010 (more than 90%) took place along two policy dimensions within two blocks of parties (Van der Meer et al. under review): a left-wing block that consists of Labour (PvdA), the Socialist Party, and GreenLeft, and a right-wing block that consists of the liberal-conservatives (VVD), the Christian-democrats (CDA) and the populist Freedom Party (PVV) and its short-lived alternative TON. There is very little exchange between these party blocks. The social-liberals (D66) function as the sole lynchpin between these blocks: D66 attracts voters from both blocks, although voters do not use D66 to bridge the gap between two blocks. There is one exception to the central electoral position of D66 in the Dutch party system: D66 hardly exchanges voters with PVV. Using this information, we constructed two additional volatility measures. Intra-block volatility covers all voters that changed their vote intention at least once within one of the party blocks mentioned above. A switch to or from D66 counts as intra-block volatility for all voters except for PVV voters. Inter-block volatility, by contrast, covers all voters that switched at least once from one party block to another (including those who switched from D66 to PVV or vice versa). It does not matter whether those switches were direct, or intermediated by a third party: we classify respondents by their preferences over the whole period. Crucial control factors Evidently, all these measures of electoral volatility depend on the number of waves respondents participated in (the more often one participated, the more likely it is to change vote intention at least once) and on the time span during which one participated (the more time passed, the more likely it is that events came to pass to change vote intention). Additionally, there are two other methodological aspects in our measures of volatility that we need to eliminate. Because we consider neither respondents who say they do not know which party they intend to vote for, nor respondents who refuse to answer to the question as respondents with substantive answers, they are less likely to be considered volatile. The more often one says not to know which party to vote for, the lower the potential score of volatility. To control for these methodological artefacts, we include these four characteristics (number of waves, time span, frequency of answering don t know and frequency of refusing to answer) as independent variables in our multivariate models of electoral volatility Independent variables To test the various models of volatility described in the theoretical section, we include several independent variables in our analysis. The socialization model is tested by the inclusion of age. Age is measured in five age groups, based on the age at which respondents first participated. Since we follow respondents over more than three years, age groups overlap (although all respondents are categorized uniquely); a person in age group 20 to 24 at the start of the interview waves in 2007 can be 26 years old in February To test the cognitive sophistication model we include level of education. Educational level is measured in eight ordered categories, ranging from primary education (1) to academic education (8). We include the same variable also squared to test the hypothesized non-linear relationship. 11 To investigate whether social embeddedness has an influence on volatility, the effect of religion, children, household composition, marital status, involvement in a sport association, a trade union and being a 11 Evidently, we would have wanted to include more direct measures of political sophistication. Unfortunately, those were not available for the whole data set. In March 2011 additional data were collected for a subset of the sample. Those data covered measures of political interest, involvement and (self-reported) knowledge as indicators of the cognitive sophistication model. Additional analyses using these measures showed further confirmation for the cognitive sophistication model. While knowledge and interest are important determinants of electoral volatility (they hamper volatility), they also explain the original relationship between level of education and volatility. As such, there is evidence for the causal pathways suggested by the cognitive sophistication model.

9 member of a political party are also tested. We operationalized religion by denomination, using dummy variables for the nonreligious, Protestants, Catholics, Muslims, and an other category for the remaining religions. Having a child is a dummy variable in which 1 means that the respondent has at least one child. Household composition is measured in five categories: living with partner, single parent, living with parents, single, and an other category. Marital status includes the categories married, divorced, widowed, and not married. The last three variables of the social embeddedness model are dummy variables for involvement in a sport association, involvement in a trade-union, and whether someone is a member of a political party. For the socio-economic vulnerability model, income, daily activity, sector, and house ownership are included. Income is measured in seven ordered categories, ranging from no income to twice the standard income. To allow for the expected non-linear relationship, the square of income is also included. Daily activity has six categories: unemployed, houseman/wife, school, pensioned, other, and employed. Sector of employment is categorized as the six most common branches of industry: agriculture, industry, commerce/finance, services, research/education, and government. All other possibilities are subsumed in an other category. House ownership is included as a dummy variable. To investigate whether media use affects volatility, dummy variables for the biggest Dutch newspapers (Telegraaf, AD, NRC, De Volkskrant, Trouw), regional newspapers, free newspapers, and no newspaper are included. Telegraaf is often considered the least politically sophisticated newspaper in the Netherlands, and comes closest to the image of a tabloid. NRC, De Volkskrant, and Trouw are often described as the typical highbrow newspapers in the Netherlands. Finally, gender and living environment are included as control variables. A dummy registers males as 1 and females as 0. To test whether living environment has an effect on volatility, the possibilities are split up into three dummy variables: big city, including the 11 biggest cities of the Netherlands, small city, and village Methods of analysis We test our hypotheses on dichotomous measures of volatility as well as on count measures. We therefore apply respectively binomial logistic regression models and Poisson regression models in our analysis. 4. Results: mapping gross electoral volatility 4.1. Shifting vote intentions Figure 2 displays the electoral volatility between subsequent opinion poll waves from November 2006 through June The bottom (black) line in the figure shows the net volatility, i.e., the percentage by which party support changed compared to the previous opinion poll wave (similar to Figure 1 for elections). It thereby compares the extent to which the net outcomes would be different across these waves. The net volatility levels are rather large. 12 The average net volatility in the 51 polls on the full panel is about 4 per cent. 13 Of course, volatility is likely to be higher when there is a larger time interval between two subsequent waves; by contrast, vote intentions will change less if there are many waves in a short time span. Regardless, we see that the relative share of support for the political parties in our data set changes continuously. Net volatility peaked at several moments. The largest peak was produced in the wave of November 12, This wave directly followed the split of MP Rita Verdonk from VVD to erect her own political party, Trots op Nederland (Proud of the Netherlands). The new party was included in the wave as a potential vote intention. The split of - then popular - Verdonk led to major shifts in vote intentions, mainly at the right side of the political spectrum. In total, the net outcome of the opinion panel changed with 15% compared to the previous wave. A second, smaller, peak took place in the first week of October It was mainly caused by the takeover of the ABN Amro bank by the Dutch government, directly after the bank crisis had hit. Vice-PM and Labour Party leader Wouter Bos benefited of a sudden image of decisiveness: his party had a sudden surge in the opinion polls, harming mainly voters of SP and TON. 12 Net volatility between waves even exceeded the net volatility between election outcomes. If elections had been held in April 2008, for instance, net volatility would have been 29.7% - higher than net volatility was at the June 2010 elections. 13 Of the 58 polls, 7 were held among a specific subsample of the panel. These polls are not included in the calculation of net or gross volatility in Figure 2.

10 Volatility increased again after the coalition broke down (February 2010), when Job Cohen replaced Bos as Labour Party leader (March 2010) and as the elections approached (May and June 2010). The level of volatility in those months is rather conservative, due to the large number of opinion polls in a short time span. Figure 2 Net volatility 1 and gross volatility 2 : shifts between 1VOP waves 1 Net volatility is the sum of the aggregated electoral growth of all winning parties in comparison to the previous opinion poll wave (Pedersen 1979). The figure includes respondents who participated only once. 2 Gross volatility is the percentage of citizens changing their vote intention compared to the previous wave they participated in (gross volatility) or compared to the immediately preceding wave (gross volatility paired). Source: 1VOP Yet, at the root of these sizeable (aggregate) net volatility rates there are even larger (individual) gross volatility rates. The share of respondents that change preferences is larger than the share of support held by parties that change ownership: opposite changes cancel each other out. The two grey lines in figure 2 represent two alternative measures to assess gross volatility. Both show the same patterns similar to the patterns in net volatility. Gross volatility is continuously larger than net volatility. The difference between the light gray and the black lines is on average more than eight percentage points. Even when net changes are minimal and support for political parties thus remains rather stable (as in January 2007, when net volatility was 1%), a sizeable group of respondents had changed their vote intentions (in the same wave, gross volatility of 5% or 8%, depending on the measure). Hence, constant net configurations tend to obscure rather large changes of voters preferences. This only underpins how volatile Dutch voters are. 4.2 Volatile voters Building on the repeated waves, we classify respondents electoral volatility in various waves: (1) dichotomously, by distinguishing which citizens changed their vote intention substantively at least once; (2) by the frequency with which citizens changed vote intention substantively; and (3) by the number of parties that respondents intended to vote for at least once. Figure 3 shows the relative frequency with which the respondents in the 1VOP changed their vote intentions (the percentages add to 100%). By far the largest group of respondents can be considered stable voters: 45%, has never changed, while 75% of the respondents changed three times or less. Although voters could potentially change vote intentions 57 times, the actual maximum was 30. Yet, the percentage stable voters is an overestimation; it is inflated by the large number of respondents who only took part in 2 waves or took part only during a short time span: respondents who participated in the 1VOP more often or during a longer time span have more options to give differential answers.

11 Figure 3. Frequency of changes in vote intentions, the Netherlands, Nov June 2010 Source: 1VOP Figure 4 breaks these figures down by the frequency of participation in the 1VOP. The figure shows that the relative frequency by which respondents change vote intention hardly differs with the number of waves in which they participated. Figure 4. Frequency of changes, by number of waves Source: 1VOP When we further study the group of citizens that did change vote intention at least once (the 55% in figure 3: respondents), we find that they change in 23% of the opinion polls (sd: 19%) in which they take part. These volatile respondents on average intended to vote on 2.75 different political parties of the potential 13 parties we distinguished during this time period (minimum 2, maximum 10; SD=0.98). 52% of the volatile voters mentioned only two political parties. The number of parties that are mentioned increases quite consistently with number of waves in which the volatile respondents participated, from the logical minimum of two parties (for those who participated in only two waves) up to three parties (for those who participated in forty waves or more).

12 Figure 5. Total number of parties mentioned, by number of waves Source: 1VOP 5. Results: explaining electoral volatility 5.1. General measures of volatility Table 1 describes the results of the multivariate regression analyses we performed to explain the three modes of volatility by the theoretical models discussed above. Although their effects are not displayed in Table 1, the models control for various methodological artefacts (frequency of participation, time span, frequency of undecided answers and frequency of refusing to answer). When we compare across the three modes of volatility, we find some determinants to have more consistent effects than others. Yet, there is only one variable for which we find outright opposite effects, namely the effect of Protestantism. Compared to being non-religious, Protestants are less likely to change vote intention. However, given that they changed at least once, they are more likely to change often. Hypothesis 1 finds no confirmation. Rather, we reach opposite conclusions. In the Netherlands, young voters (below 42) are less likely to be volatile than older voters, especially those older than 55: the young are less likely to change vote intention (b = for year olds), and even when they changed vote intentions, they are less likely to change often (b = -0.13, and -0.06) or to name many parties (b=-0.16 for year olds). Socialization thus does not seem to have any stabilizing effect. This finding is curious, as previous studies consistently found that volatility decreased with age (Schmitt-Beck et al. 2006; Söderlund 2008; Kuhn 2009). What might be the cause of our differential finding? There are some technical differences between those studies and ours: we cover a lot more panel waves, and our analyses are not restricted to actual (recalled) votes during elections. It might be the case that older voters are more consistent when it actually matters (i.e., at the voting booth), but less consistent when it does not. We will test this idea in paragraph 5.3. Alternatively, the effect could be due to the rise and decline of TON, which briefly attracted many conservative voters in 2007 and Our second model is the cognitive sophistication model. Hypothesis 2a expected that, generally, volatility should first rise with sophistication and subsequently decline again. Indeed, we find consistent evidence for a non-linear effect in that direction. Volatility is lowest for those who only had primary education (until the age of twelve) and for those who studied at university. It is highest for voters whose highest level of education is (the advanced professional or the pre-academic track in) secondary education on all measures of volatility. Hypothesis 2a is therefore fully supported by the models in table 1. We find less consistent support for the social embeddedness model. We had expected people without family ties, religious ties, and associational ties to be the most volatile. Yet, with regards to family most effects are non-significant, and the significant effects point in the opposite direction. People living with a partner are more likely to change vote intention at least once than singles (b=0.103). Not the unmarried but the widowed and the divorced are most likely to be volatile; whereas unmarried and married voters do

13 not differ significantly. Similarly, religiosity is not a stabilizing characteristic either. Compared to being non-religious, Protestants are less likely to change at least once, Catholics are no different and Islamic voters and other religious voters are in fact more likely to change once. And given changing vote intentions, Protestants and Catholics are more likely to change frequently and more likely to mention many parties. In any regard, there is little evidence for a stabilizing effect of religiosity. 14 Finally, we find that membership of some types of association does diminish electoral volatility. However, it is hardly surprising that membership of a trade union and especially membership of a political party have such effects: voters in these organizations already have strong institutional ties to parties that will limit their likelihood to change. By contrast, membership of sports associations (many still a product of pillarization) does not dampen volatility but stimulates its frequency (b=0.21). Generally, hypothesis 3 finds little support; solely voters who are involved in a trade union or political party are less volatile than the uninvolved. The fourth model emphasized socio-economic vulnerability. There is little evidence that vulnerable voters (low income, unemployed) are more likely to be volatile. None of the significant effects points in that direction. We therefore reject hypothesis 4a. Rather, we find strong support for hypothesis 4b, which argued that the effect of income should be curvelinear. Indeed, according to the models in Table 1 the most volatile voters are those with a median income (4). The likelihood to change vote intention at least once increases with income until the median income group, and decreases afterwards. The same goes for the total number of parties that were mentioned. Our final model argues that use of media in general, or else some types of media, stimulate electoral volatility. We find no consistent evidence that reading a newspaper per se stimulates volatility. Effects of newspaper reading go in opposite directions. We find more support for the idea that the type of media matters. Volatility is highest amongst the readers of the most popular Dutch tabloid, Telegraaf. Readers of that newspaper score highest on all measures of volatility. Yet, reading highbrow newspapers (such as Volkskrant, NRC and Trouw) does not obstruct volatility; effects are scarce and inconsistent. All in all, hypothesis 5b finds some support. Table 1: Regression models of electoral volatility (unstandardized coefficients) Volatility Frequency of (dichotomous) 1 change 2 Total number of parties 2 Demographic control Gender (ref: female) Male.043 (.022).013 (.010).016 (.005) Living environment (ref: village) Big city.085 (.027).025 (.012).007 (.006) Small city.048 (.022).008 (.010).003 (.005) Socialization Age (ref: 40-57) (.061) (.029) (.012) (.030) (.014) (.006) (.030).074 (.013).024 (.006) (.044).115 (.019).022 (.009) Cognitive sophistication Level of Education.096 (.034).030 (.015).022 (.007) LoE (Squared) (.003) (.002) (.001) Social embeddedness Having a child (.028).023 (.013).002 (.006) Household composition (ref: single) Living with partner.103 (.040).039 (.018).010 (.008) Single parent (.063).087 (.028).022 (.013) Living with parents.020 (.070).020 (.035).013 (.015) Other.003 (.050).045 (.024).017 (.011) Marital status (ref: not married) Married (.040).023 (.018).014 (.008) Divorced.156 (.042) (.019) (.009) Widowed.073 (.061).071 (.026).028 (.013) 14 This might have to do with the fading of the religious cleavage in Dutch politics, which gives religious voters a less natural position in the party system.