Nationwide localism
Voter-candidate proximity effects in two types of Dutch elections
Mark Hessels Student no. 5936055
Master Thesis
Supervisor: T.W.G. (Tom) van der Meer Second reader: A. (Andrea) Ruggeri
University of Amsterdam
Research Master in Social Sciences Empirical Analytical Track
Specialization: Comparative Politics
2 Abstract
This study investigates geographic voting behaviour in the Netherlands. This country is an interesting case to conduct this geographic voting behaviour, since there is a lack of institutional features ensuring regional voting behaviour in the Netherlands. Using the votes for all candidates in the most recent national election as well as the most recent provincial elections, this study tests in two ways the extent of local voting behaviour in the Netherlands. Based on a shared administrative community (a candidate MP lives in the municipality of the voter), candidates get a significantly higher vote share in their own municipality of living, and also in their region and province of living. It is also shown that candidates get a lower vote share when the distance between candidate and voter municipality is higher.
3
1. Introduction
The Christian Democrats (CDA) obtained 13 seats during the most recent Dutch general election (September 12, 2012). Four candidate Members of Parliament from that party were elected through obtaining sufficient preferential votes (numbers one, two, three and number 39 on the ballot). Of the remaining nine seats, the numbers four through twelve on the ballot were elected. Number 39 on the ballot is incumbent MP Pieter Omtzigt. After a regional campaign, Omtzigt has been re-elected in the House of Representatives because of the regional votes. In Twente, the region in which Omtzigt lives, 52% of the people who voted for the CDA voted on Omtzigt. Just these votes were enough for Omtzigt to obtain a seat exclusively through preferential votes.
This is an example of regional voting in a national election: people cast a preferential vote on a candidate living in their own region. Regional voting in a national election is a form of descriptive representation, which relates to voter-candidate correspondence: one or more shared physical and/or other characteristics between voter and candidate, which influences voting behaviour. As the example shows, during the last elections the majority of people, who voted for the CDA and live in the region of Twente, voted on a candidate that lives in their region. As a result, the voters changed the list order and ensured that the number 39 on the list was elected (instead of the number 13 on the list).
In the Netherlands it is possible to vote on the same candidate in the whole country1, as the whole country consists of only one voting district. Furthermore, it does not matter on which candidate somebody votes, since every vote is added to the respective political party’s total. Consequently the list of prospective Members of Parliament is not region-specific, but nation-wide, and most national media attention focuses on the leading candidates of the (main) political parties, who have to appeal to the whole nation, and not a single region. Regional media outlets could make other considerations and can focus more on regional candidates. Regional voting only matters when a candidate gets more than 25% of the electoral quota. This means that a candidate should have more than 15.708 votes during the last elections to be elected by preferential votes.
This thesis attempts to answer the question under which circumstances locality matters (living in the same municipality, region and/or province). Although significant research has been conducted on preferential voting (for a overview, see Karvonen (2004)) and descriptive linkages between candidates and voters (Cutler, 2002), less research has been conducted on voting behaviour with respect to a candidate living in one’s own locality. In the latter case, most research is about electoral systems with institutional features ensuring regional voting behaviour (Bowler et al., 1993;
1 For every candidate, the name, gender and place of residence are listed on the ballot (apart from the name of
the political party and the position on the list). There are even candidates with a foreign place of residence on the ballot. Not only the place of residence is listed in these cases. In addition, the abbreviation of the country is listed on the ballot, for instance ‘London, UK’.
4 Cutler, 2002; Gimpel et al., 2008; Gorecki & Marsh, 2012; Key, 1949). Little is known about the reasons for a vote on a candidate from the voter’s region in central PR-list systems. The research question is therefore:
Under which circumstances do people vote on a candidate from their own region?
Data from the most recent Dutch national and Dutch provincial elections will be used to test the conditions for a ‘local’ vote. The Netherlands is selected as a case study, since it is the least-likely case. There are no institutional features ensuring regional voting behaviour. The emphasis is on the various parties, not on the party lists.
The thesis is structured as follows. First the contemporary literature on localism in voting, as well as theories on descriptive representation and preferential voting will be analysed. Resulting from this, eight hypotheses will be presented for investigation. In the third section, the data and the case-selection will be explicated. The fourth section shows the empirical results, which shows that candidates get significantly a higher vote share in their locality of living. The conclusion and discussion are section five.
5
2. Theory and expectations
2.1 Descriptive representation
The subject of descriptive representation is part of the wider research area on political representation. Descriptive representation or the “politics of presence” (Phillips, 1995) means that “representatives are in their own persons and lives in some sense typical of the larger class of persons whom they represent” (Mansbridge, 1999: 629). Most attention is paid to ‘visible characteristics’ with respect to descriptive representation (for instance, sex and race). However, ‘shared experience’ is also a form of descriptive representation (for instance, living in the same region) (Mansbridge, 1999). These two categories are not mutually exclusive; a female voter can decide to vote on a female candidate, who is from her own region as well.
Mansbridge notes the significance of descriptive representation:
“Seeing proportional numbers of members of their group exercising the responsibility of ruling with full status in the legislature can enhance de facto legitimacy by making citizens, and particularly members of historically underrepresented groups, feel as if they themselves were present in the deliberations […] Seeing women from the U.S. House of Representatives storming the steps of the Senate, for example, made some women feel actively represented in ways that a photograph of male legislators could never have done” (Mansbridge, 1999: 650-651)2.
Mansbridge asserts that people feel better represented if the elected MPs are more similar to themselves (Mansbridge, 1999). Another reason for people to vote on a candidate with shared characteristics is that people might assume that that candidate will act in the same way as the voter would do, because of their shared characteristics:
“Demographic facts provide a low-information shortcut to estimating a candidate’s policy preferences […] Characteristics such as a candidate’s race, ethnicity, religion, gender, and local ties […] are important cues because the voter observes the relationship between these traits and real-life behavior as part of his daily experience. When these characteristics are closely aligned with the interests of the voter, they provide a basis for reasonable, accessible, and economical estimates of candidate behavior” (Popkin, 1994: 63-64).
2
In response to particularly normative theorists who argue against descriptive representation, Mansbridge acknowledges that:
“The primary function of representative democracy is to represent the substantive interests of the represented through both deliberation and aggregation. Descriptive representation should be judged primarily on this criterion. When nondescriptive representatives have, for various reasons, greater ability to represent the substantive interests of their constituents, this is a major argument against descriptive representation”
6 Not in all countries is it for people possible to vote on a certain candidate of a political party during the elections3. This means that people in these countries should accept the ranking of candidates drawn up by the political party. In other countries, it is possible to vote on one candidate and if that candidate gets enough votes, people can change the order of candidates on the ballot. The literature review in this thesis will ignore countries in which preferential voting is impossible, due to the reason that people have then no option to vote (or vote not) on a candidate with certain shared characteristics during the elections. In this thesis, preferential voting is defined as “the opportunity to choose among several candidates of the same party” (Karvonen, 2004: 203)4. If voters could vote on a specific candidate, they could use shared characteristics as a means to decide on which candidate the voter should vote. One type of descriptive representation is localism in voting, which will be elaborated on in the next couple of paragraphs.
2.2 Localism in voting
Localism in voting is certainly not a new subject, as Key posited the ‘friends and neighbours hypothesis’ as early as in 1949. The theory is that the likelihood of a vote for a candidate is higher for voters that live in the same locality (for instance a county) as a candidate, ceteris paribus. The challenge with this topic is that almost all research has been conducted in the United States and, the concept (localism in voting) is far from thoroughly problematized; the conditions that are essential for localism in voting is unclear. Below four problems with the existing literature on regional voting are outlined.
2.2.1 Shared administrative communities versus geographical proximity
The first problem is that there are two types of studies about this localism in voting. The first group of studies looks whether a candidate lives in the same locality as the voter (Black & Black, 1973; Kjar & Labrand, 2002; Tatalovich, 1975); hence these studies look at shared administrative communities. The other group of studies use another starting point: geographical proximity. The hypothesis is that geographical distance between voter and candidate is negatively correlated with the likelihood of vote. The closer a candidate lives to a voter, the higher likelihood of a vote on that candidate from that voter, ceteris paribus (Bowler et al., 1993; Cutler, 2002; Gimpel et al., 2008; Górecki & Marsh, 2012; Van Holsteyn & Andeweg, 2012).
For the American case, Key asserted that the candidates for state office have a home county bonus (Key, 1949). In Australia, testing personalization of politics at the constituency level, “it
3
In most countries, party members could influence the sequence of candidates before the elections. That is however not the focus of this thesis.
7 remains the case that the Australian electoral system – and the choices that voters must make at the ballot – remain predominantly party choices” (McAllister, 2013: 7).
In the general election of single-member districts in England, a recent study shows that “local is better” (Arzheimer & Evans, 2012: 309). This study uses geographical proximity. It was found that “the distance between voter and candidates from the three main parties (Conservative, Labour and Liberal Democrat) matters in English constituencies, even when controlling for strong predictors of vote choice, such as party feeling and incumbency advantage” (Arzheimer & Evans, 2012: 301).
For the Netherlands, geographical proximity has the weakest effect of all variables (gender of the candidate, ethnicity, position on the ballot, incumbency) in a study about preferential voting (Van Holsteyn & Andeweg, 2012). However, the researchers did not test whether the likelihood of a vote on a candidate is higher if the candidate lives in the same locality as the voter. In Ireland, with its single transferable vote (STV) electoral system, the likelihood increases when a candidate lives closer to a voter (Górecki & Marsh, 2012).
In the case of England, candidates could choose to have their full home address on the ballot or only the constituency of residence. For the candidates with the full information on the ballot, the researchers were able to estimate distances between voters and candidates in a detailed manner. This study concluded that “[t]he next step [...] is to refine the definition of ‘local’. [...] Place of birth, regional identity and other dimension of localness all matter. Some of these are potentially, if arduously, quantifiable, and may indeed matter more than geographical distance” (Arzheimer & Evans, 2012: 309).
2.2.2 Shared place of living versus the candidate’s place of birth and the voter’s place of living The second problem is the local link. It is argued that people who want a local representative mean that they want a MP “who will, in some form, come from that area” (Childs & Cowley, 2011: 4). One form of coming from that area is living in that area currently. Another form is to be born in that area:
“Long-term residents in a town often argue for electing to office someone born in the town on the implicit grounds that lifetime experience increases the representative’s common experiences with and attachment to the interests of the constituents. [...] “Being one of us” is assumed to promote loyalty to “our” interests” (Mansbridge, 1999: 629).
This could be a reason for women to vote on a female candidate, for instance, or – and that is the main focus of this thesis – people could vote on a candidate who lives close to their home as that candidate shares the characteristic of living in the same locality as the voter.
Place of birth as form of a local tie is neglected in most studies (Górecki & Marsh, 2012; Van Holsteyn & Andeweg, 2012), since “not easily quantified” (Arzheimer & Evans, 2012: 303). Other studies mention the possibility, without testing it (Gimpel et al., 2008; McAllister, 2013). Studies that
8 combine birthplace and place of residence look only at the party behaviour of selecting candidate-MPs on the ballot (Pedersen et al., 2007; Latner & McGann, 2005)5 or write only a normative story with arguments in favour of descriptive representation (Childs & Cowley, 2011).
2.2.3 Regional identity
The third problem relates to the different cultural identities between people across a nation and its possible influence on localism in voting. Regional identity is a “dimension of localness” (Arzheimer & Evans, 2012: 309). In the United States, for example:
“Not all states and regions are characterized by uniform levels of local loyalty, so variation in this factor should interact with geographic proximity to create variation in home state or regional advantages. For instance, voters from the Southern or Western states are often characterized as having higher levels of state or regional loyalty arising out of common identification grounded in historical experience. Hence, one might expect citizens from these two regions to be more likely than citizens of other regions to support presidential tickets with local geographic attributes” (Garand, 1988: 90)6.
It is expected that local loyalty interacts with geographic proximity. Another way of looking at differences in regional identities that cause different graduations of local loyalty is to look at the proximity of “the centre of politics” (Bengtsson & Wass, 2011: 151). In most countries, most national decision-making institutions, such as the national parliament, are located in the same city. For Finland, it is tested whether people living far away from the capital (Helsinki) have “a stronger preference for geographical representation” (Bengtsson & Wass, 2011: 151). The authors found a significant effect in the expected direction. However, they do not explain the mechanism behind expected differences between people living close and people living far away from the centre of politics. It could be that the media are also situated in the capital and that there is possibly a bias for news about the capital and its surroundings. Places that are a long distance away from the capital might get less media coverage. Perhaps politicians do talk less about these places and therefore people from these places prefer to vote on candidates from their own locality.
5
Geographical representation was studied for the Netherlands and Israel (Latner & McGann, 2005). For the Netherlands, they concluded on the base of the Dutch general election of 2003, that the Dutch Parliament is geographical representative. However, they showed only descriptive information (comparing the number of elected MPs per province to the relative share of voters from each province) and conducted this information only at the provincial-level. Moreover, they neglected voting behaviour. Instead, they only looked at the birthplace and place of residence of the elected MPs.
6
In this study, it has been found that “regional loyalty […] increases GOP support by 1.77 percent (for presidential candidates from the West)” (Garand, 1988: 99). This effect was not found for Republican presidential candidates living in the South. For the Democratic presidential candidate on the other hand, a positive effect of living in the South was found and not in the West (Garand, 1988).
9 Furthermore, in a study about the Netherlands and Israel, it is hypothesized “that regional identity is stronger and more salient in peripheral regions” (Latner & McGann, 2005: 713), “either because there is a distinctive culture (Friesland, for example, has its own language) or because it is intrinsically important” (Latner & McGann, 2005: 718). Although these authors argue that people from peripheral regions could find it more important to vote on a candidate from their own region compared to people from urban regions, they did not test this assumption. In this thesis, however, it will be tested whether there are regional (cultural) differences in voting on a local candidate.
2.2.4 Election types
The last not well problematized component is differences of localism in voting behaviour between various elections in a country. Most studies do not compare their results of one election with other elections in the same country. Hence, they very often use only one election (however sometimes a series of elections of the same type).
Local elections are often indicated as ‘second-order elections’ (Reif and Schmitt, 1980). This means that these elections are less salient for people and therefore relatively less people actually vote in these elections (lower voter turnout). The voting behaviour of people is different compared to national elections as ideology and policy issues are of less importance (Reif and Schmitt, 1980; Norris and Reif, 1997).
Most research analyses national representative bodies. Comparing presidential candidates, Kaufman (2004: 20) argues that the pool of candidates for local elections is more diverse. This makes it more easy to vote on a person that shares one or more (for the voter important) characteristics with the respective voter.
In the next section, four hypotheses will be derived from these four, not well problematized components, of localism in voting7. In doing so, the typical research problems on regional voting are overcome.
7 Another aspect of localism in voting is the number of candidates having a connection to a certain locality. If
there are more people for instance living in the same locality from one party, do all these candidates get more votes in that locality compared to all other localities in which they do not live, ceteris paribus. Johnston calls this possibility “the density of candidates” (Johnston, 1974: 423). Most studies do not include this, since these studies are about single-districts with one candidate from each political party on the list. However, in district magnitudes with two or more places for MPs, it becomes interesting whether there is still a positive effect of living in the same locality if another candidate from the same party is also living in that locality. Some studies control for the number of candidates in a constituency (Gorecki & Marsh, 2012: 570-572 (Ireland); McAllister, 2013: 3 (Australia)). In Ireland, it was found that if the number of other candidates from the same party increases, the probability of a candidate from that party to be selected decreases, ceteris paribus. This thesis will control for the density of candidates.
10
2.3 Hypotheses
2.3.1 Shared locality of living
As this thesis analyses which circumstances are important for local voting behaviour, the first test should be whether there is an effect at all. Therefore the supposition that localism in voting is significant in the Netherlands will be tested in two ways: shared administrative communities and geographical proximity. In the literature, no studies were found who used both methods: all studies use only one form of the theory. When both types are compared and both types are showing the same results, then the results are more robust.
Three administrative communities are used in this thesis, from low to high: the municipality, the region and the province8. This means that there could be a shared administrative community at three levels. It is expected that a shared municipality of living is more positive significant compared to only a shared region of living (corrected for one of the municipalities in that region in which the candidate lives). However, a shared region of living is expected to be more positive significant compared to only a shared province of living (corrected for municipality and region of living). Therefore, the first hypothesis is:
The lower the level of a shared administrative community, the more positive the significant effect on the candidate’s vote share compared to all other localities (H1: Shared locality of living)
2.3.2 Geographical proximity of living
The second hypothesis regarding geographical proximity as a factor is the same as for all other studies that use this approach:
The closer a candidate lives to a voter, the higher likelihood of a vote on that candidate from that voter, ceteris paribus (H2: Geographical proximity of living)
2.3.3 Connection of candidate’s locality of birth & voter’s locality of living
In the theory section, it is mentioned that some studies use the place of residence of a candidate and others use the place of birth of a candidate. This thesis will look upon both these aspects of localism. The first part was tested in the first two hypotheses, now the place of birth of a candidate will be tested in the next two hypotheses.
A connection between the place of birth of a candidate and the place of living of a voter means that it is unknown whether the voter is born in the same locality as the candidate MP. It is only concerned with whether the candidate was born in place A and that the voter lives during the
8
The municipality and province are administrative communities in which specific elections are hold. Region is a ‘cultural’ division. For the sake of clarity will all three levels be named as localities and administrative
11 elections in place A. It could be that the place of residence of a candidate is the same as the place of living of that candidate, in that case it is impossible to see which is more important. In all the other cases, it could be found out if place of birth or place of residence is more important for localism in voting behaviour. For place of birth, the same hypotheses as for place of living are formulated with the same argumentation:
The positive effect of the candidate’s birthplace on his vote share is the highest in the lowest level of the administrative community that the birthplace falls in, (H3: Connection candidate’s locality of birth & voter’s locality of living)
2.3.4 Geographical proximity between the candidate’s place of birth & the voter’s place of living The hypothesized effect from hypothesis 3 could also be tested by means of the proximity method:
The closer a candidates place of birth is to the place of living of a voter, the higher the likelihood of a vote on that candidate from a voter from that place of living (H4: Geographical proximity between the candidate’s place of birth & voter’s place of living)
2.3.5 Randstad versus rest of the country
In the theory section on regional identity, the first difference mentioned was the different levels of local loyalty. However, in this thesis it is not possible to test this, as there is no information about the voter, apart from the assumption9 that the voter lives in the municipality in which he or she votes.
The second difference is the distance from the political capital of a country. A rough difference in regional identity could be the difference between the Randstad and the rest of the Netherlands. The Randstad is the most urbanized area of the Netherlands (a conurbation with the four largest Dutch cities including the political capital and its surroundings) and the rest of Netherlands is more rural. The authorities and for instance most national media outlets are located likewise in the Randstad. Finally, most candidate-MPs are living in the Randstad. The hypothesis is therefore:
The beneficial effect for a candidate not living in the Randstad is such, that it will get more votes in its own locality of residence compared to candidate MPs that live in the Randstad (H5: Randstad versus the rest)
9
All eligible voters get a voter card to vote in the municipality of living. If someone wants to vote in another municipality, then he or she should request that. Therefore, the assumption is that most votes are from people living in the municipality where they vote. There are no exact figures known how many people do a request to vote somewhere else. If there is an effect found of living in the same municipality, it is then an underestimated effect, since the link between the people that could vote on a candidate from their own locality (but vote somewhere else) is not visible.
12 2.3.6 Proximity to the centre of politics
The fifth hypothesis could also be tested by means of the distance between the centre of politics and the place of living of a voter. This is similar to the test of Bengtsson & Wass (2011). Instead of comparing the Randstad to the rest of the Netherlands, the distance from The Hague (the seat of the Dutch government) to each other municipality could be calculated and tested:
The greater the distance between the place of residence of a candidate MP and the centre of politics, the relatively more votes a candidate MP gets in his locality of living (H6: Proximity to the centre of politics)
2.3.7 First- and second-order elections comparison
The last problematized part in the theory section was with respect to the differences between election types. It is assumed that people know less in non-national elections and vote differently (Reif and Schmitt, 1980; Norris and Reif, 1997). As people know less in second-order elections, it could be that people vote relatively more often on a local candidate compared to national elections. On the other hand, in most countries, regional elections consist of only candidates living in that administrative unit which implies regional voting behaviour. This means that the region is represented, a priori. It becomes interesting to study whether there is an effect of living in a certain locality, even at this lower election level.
The lower the level of a shared administrative community, the more positive the significant effect on the candidate’s vote share compared to all other localities (H7: Shared locality of living in provincial elections)
The lower the level of an administrative community in which the candidate MP was born, the more positive the significant effect on the candidate’s vote share compared to all other localities (H8: Connection candidate’s locality of birth & voter’s locality of living in provincial elections)
2.4 The Netherlands
Although the case-selection has been introduced in the introduction, the reasons for selecting the Netherlands as case should be clear. To investigate whether regions matter even in a central party system, it is necessary to select an electoral system without institutional features which ensure regional voting behaviour. This means that there should be not more than one electoral district nationwide and people have to vote on a candidate-MP (instead of only on a specific party). The latter is because of the fact that if people could only vote on a party, it is impossible to differentiate between the various localities of the candidates from that party. The first requirement (one district nationwide) fits the Netherlands and Israel. The second requirement however is only valid for the
13 Netherlands, as people could only vote on a specific party in the elections for the Knesset in Israel (as such no specific candidate can be chosen) and Dutch voters could only vote on a specific candidate-MP. As a result, the most recent national and provincial election in the Netherlands is chosen for this thesis as data source.
14
3. Data and methods
3.1 Data
The number of votes per candidate per municipality for the general elections as well as the provincial elections is provided by the Dutch Electoral Council (Kiesraad). In addition, this organization delivered the names of the candidates, the party names, the position on the ballot, the gender of the candidate, and the residence of all the candidates. In other words, it provided all the information published at the ballot in addition to the number of votes for each candidate per municipality for both elections.
Other information that will be used for the analysis, such as the birthplace of the candidate or the age of the candidate, has been collected through official websites from the government. Place of birth, age, and term are only collected for elected MPs in both elections, because this information is not available for all 972 candidates from the Dutch general election and the numerous candidates from the Dutch Provincial Elections. The place of birth is not known from every candidate out of 972 candidates in total.
This information is fully available at www.parlement.com for the elected MPs in the general election, however for the elected candidates from the Provincial Elections it was not always possible to find all the information on official websites. From some provinces, all the information needed was published on their websites, however most provinces had little information published on their website. In that case, there was communication with the various provinces and some send the right information. In some cases, where the information was not found through using the above measures, information from non-governmental websites was consulted (for instance political party websites or personal websites of candidates). In most cases, the candidate was contacted by email with the question if they could give the author the needed information. From the 127 mails, 68 (54%) responded to this quest for information with the necessary information.
3.2 Methods
The data consists of clustering at various levels. The number of observations is the product of the number of competing candidate-MPs and the number of municipalities. As a result, there is clustering within candidates as well as within municipalities: the share of votes for a candidate in a municipality is nested within the municipalities as well as within the candidates. To correctly deal with these characteristics of the data, a multilevel design is indispensable (Snijders & Bosker, 1999), though a ‘normal’ multilevel design is theoretically insufficient. The vote share of a candidate in a municipality, the candidates, and the municipalities are not on three different levels. If that was true, then a normal multilevel design was sufficient. With this type of data, the lowest level is the vote share of a candidate in a municipality (logit); yet the candidates as well as the municipalities are at
15 the same (second) level (see figure one). To cope with these specific features of the data in the correct manner, a cross-classified multilevel estimation procedure will be estimated. As a result, the standard errors will be estimated on the right level, which was otherwise not conceivable10.
Figure 1: Cross-classified multilevel model
The cross-classified multilevel estimations were compared to a ‘normal’ multilevel estimation with only candidate at the second level, because some models were computationally not able to run. It turns out that the estimated coefficients of most variables did not change at all, and other estimated coefficients changed only maximally 0.001. Therefore, and to be able to compare, all models are estimated with only candidate at that second level (Figure 2). Hence, there will be only controlled for the unobserved characteristics of candidates.
Figure 2: Basic multilevel model
The candidate’s share in a municipality is the dependent variable that will be used in all estimations. The candidate’s share of votes on his party in a municipality is a relative measure, since it is the number of votes for a candidate relative to all the votes for the candidate’s party in a municipality:
The result of this calculation is a number between zero and one11. The dependent variable is the logit transformation of the above calculation:
10
See for information about cross-classified multilevel models the manual of Leckie (2013: 1-52) or Rabe-Hesketh and Skrondal (2012).
11
In order to avoid estimation problems, candidates with zero votes in a locality get not a vote share of zero. Hence, they get a vote share of: , which is a number very close to zero. The
Candidate’s share in a municipality + all other variables Candidates Le ve l 1 Le ve l 2
Candidate’s share in locality (logit) Dummy’s living locality / locality of birth Dummy’s locality of living, locality of birth, proximity
Candidates Candidate characteristics Dummy’s political parties Municipalities Municipality characteristics Le ve l 1 Le ve l 2
16 .
The use of a logit transformation of the dependent variable has a couple of advantages. As the dependent variable is a probability, it can only take values between zero and one. Because of this, a multilevel estimation is not valid, as it assumes that the standard errors variable can take every possible value. When the dependent variable can only take values between zero and one, then the standard errors cannot take every possible value. The transformed dependent variable does have this characteristic, so the estimate is unbiased. To interpret the estimated coefficients, these have to be transformed back to the regular probabilities. The formula for this inverse logit transformation is:
.
The choice for a relative measure instead of using the absolute number of votes for each candidate in each municipality has mainly two reasons. First, the number of votes in total varies hugely between the 415 municipalities, which have to be corrected otherwise. There are minimally 1.067 votes cast in a municipality (Rozendaal) and maximally 389.276 votes in another municipality (Amsterdam). It is on a different scale. The value of the constant is then totally dependent on the size of the municipality.
Second, the number of votes varies hugely between parties. As it is essential to see whether voters are different between political parties with respect to regional voting behaviour, the party size should then also be corrected. If you take the absolute number of votes, then votes in Amsterdam are much more important than votes in Rozendaal. It could be that a candidate from Rozendaal gets in absolute numbers more votes in Amsterdam compared to Rozendaal, which could mean relatively very few votes in Amsterdam and relatively many votes in Rozendaal.
3.3 Variables
The main focus of this thesis is the question whether candidates will get relatively more votes in the candidate’s locality of living, compared to other areas. This will be investigated through eight hypotheses. The variables of interest will be explained in the next couple of paragraphs. The operationalization of the base regressors that will be used in each model (mainly control variables) are explained in footnotes. The descriptive statistics of all these variables are listed in Appendix A.
same procedure applies to candidates with a 100% vote share in a locality, these candidates in these localities get a vote share of: , which is a number very close to one. The reason for this is that it is not possible to use logit with 0.000 and 1.000.
17 Locality of living is tested at three levels. Therefore, there are three dummy variables whether the candidate lives in the municipality, the region, and the province of the voter (yes or no). The residence of each candidate is explicitly stated on the ballot. This information has been appended to the corresponding municipalities and from all municipalities the related region and province is known. In 2011, there are 418 municipalities and in 2012 there are 415 municipalities in the Netherlands12. There are on the basis of the COROP-division (NUTS 3) 40 regions in the Netherlands and there are twelve provinces in the Netherlands. Every region belongs exclusively to one province13.
The locality of birth are again three dummy variables whether the locality of birth is the same as the locality of living of a voter14. Furthermore, information about the ethnicity has been collected to see which candidate-MPs are non-Dutch1516.
For geographical proximity it was necessary to estimate the distance between all municipalities17. Therefore, the postal code of each city hall was collected through an official governmental website (Website Overheid.nl). This information is then linked to a dataset with all distances between all postal codes in the Netherlands. The natural logarithm of this measure is taken, because it is expected that it matters that a city is 15 kilometres away instead of 30 kilometres. Yet, 315 kilometres away instead of 300 kilometres is absolutely an even large difference, however relatively not. For the distance between the place of birth of a candidate and the place of living of a voter, the exact same procedure has been followed. For the distance between The Hague and the municipality of the voter, the postal code of the Dutch House of Representatives is linked to all postal codes of again the city halls in the Netherlands. Besides, an interaction effect between this distance and the indicator that shows whether the candidate lives in a certain municipality is estimated to see the effect for a candidate MP.
12 People from three special municipalities in the Caribbean also participated in the Dutch general elections of
2012. The municipalities Bonaire, Sint Eustatius, and Saba are excluded due to a low voting turnout.
13
There are also candidates living abroad, these occur in the dummy variable ‘lives not in the election unit’.
14 The candidates that were born abroad are inserted in the dummy variable ‘not born in the election unit’. 15 The family name of a candidate could be a shortcut too (Van Holsteyn & Andeweg, 2012), since almost all of
these candidates have non-Dutch family names on the ballot. If a candidate has a Turkish family name, it could be a shortcut for Turkish-Dutch to vote on that candidate. As this form of voter-candidate linkages is not the main focus of this thesis, there are no hypotheses about it.
16
This is a different variable compared to whether the candidate was born outside of the Netherlands. In the non-Dutch candidate-variable, there are also candidates that were born actually in the Netherlands, but are considered as non-Dutch candidates (correlation between these two variables is 41.38%).
17
Of the votes in a municipality, there is no connection to an individual voter. This means that the exact address of a voter is unknown. Therefore, from each municipality, the postal code of the City Hall is used to calculate the proximity between the voter and the candidate.
18 Whether the voter lives in the Randstad or not is estimated through two variables. The first is a dummy variable of the conurbation the Randstad18. The second is an interaction effect of this variable and the indicator that shows whether the candidate lives in a certain municipality.
This thesis controls for the number of candidates in a certain municipality in two ways. At first, a candidate can be the only candidate on his party’s list that lives in a certain municipality. But there can also be more members of his party’s list that live in the same municipality. In that case, the estimated coefficient of the indicator that shows whether the candidate lives in a certain municipality will probably be lower. So therefore, different dummies are made that indicate the number of candidates from the same party living in a municipality.
In addition, a candidate will probably get less votes in a municipality if there are several candidates living in this municipality on his party’s list. For example, a candidate from Groningen will get less votes in Amsterdam, the more other candidates from his party there are from Amsterdam. So different dummies are made to correct for this effect too.
Regressors that are used as control variables are: gender (female as reference category), age and age-squared19 of the candidate; highest male/female on the ballot; position on the ballot20; experience21; interaction terms for the position on a specific party list22.
At the municipal level, two control variables will be added. The first one is the share of non-Dutch people and the second one is the share of higher educated people in a certain municipality23. This is to control for voter characteristics that could differ between municipalities. Since the votes are not linked to specific voters (we only assume that they live in the municipality where they vote),
18 The Randstad-region consists of all regions from the provinces Flevoland, South Holland and Utrecht in
addition to five out of seven in North-Holland: IJmond, Haarlem agglomeration, Zaanstreek, Greater Amsterdam, Gooi and Vechtstreek.
19
As it is not expected that there is a linear relationship between the age and the number of votes for a candidate.
20
The place on the list is divided in groups: candidate 2 & 3, 4 & 5, 6 & 7, 8 & 9, 10-14, 15-19, 20-24, 25-29, 30-34, 35-39, 40-44, 45-49, 50-54, 55-59, 60-64, 65-69, 70 and higher, with the number one as basis.
21 Experience means the number of days between the first day in Parliament or Government and Election Day.
This means that most candidate-MPs have an experience of zero days, as they have not been a Member of Parliament or a Member of the Government before. The candidate-MPs with experience are divided into three groups, minimally one day until three years, three years until six years, and six years or more experience. The reference category is as a result zero days experience.
22
In the descriptive information-section, it will be shown that the average share of the elected number one candidates is higher than 80%. This means that this information should be taken into account in the right way. As a result, a couple of interaction terms are added to the base model: the number one candidate of a party times his or her party, and the rest of the candidates times their party, with the leading candidate of the largest party as reference category (VVD – Rutte). While estimating all the models, it has been found that with this approach, the differences between the number one candidates versus the other candidates can be best predicted.
23
Economic measures at the municipal level were not available (this was only available at the regional level, hence not specific enough)
19 these two variables are not analysed in detail. These two variables are interacted with a couple of variables that are candidate specific, in order to control for candidate-municipality effects24.
3.4 Estimation restrictions
This thesis will use models with only elected candidates as well as models with as much candidates as possible. This has to do with data restrictions and estimation restrictions. The data restrictions have already been discussed in the data section. The estimation restrictions are the problem that it is computationally not possible to estimating the models with all candidates and a the full set of regressors. As will be shown later on in the results, a model with all candidates for the general elections is possible if the only regressors are about the locality of living. If more variables are added, the computers cannot cope with this. Therefore, many models have been estimated to see whether there are different effects. Since model after model, the estimated coefficients of the key variables are approximately similar, it is not very problematic. Due to the data restrictions, the decision has been made to show almost only models with only the elected candidate MPs for the general elections, to be able to compare all the models properly.
Second, if we use only the elected candidates, then it is also not possible to add variables at the candidate-level (hence, candidate-specific variables that are in each municipality the same). Therefore, all variables are added at the lowest level. There are models estimated with consistently one variable at the candidate-level which show that the results are very similar.
Another issue is that the provincial elections are actually twelve different elections holding at the same day, as there are twelve provinces in the Netherlands. Logically, in this election almost all candidates live in the same province and people can only vote for candidates living in their province. In the national elections in the Netherlands, it is of course possible to vote on a candidate that is not living in your province. Also, party lists are shorter for provincial elections, which means that the average share of votes of candidates is larger. So these are two structural differences in these elections, and it is not possible to correct for these structural differences and therefore to compare these 12 provincial elections with one national election with this type of data. The case in which there are three levels, with votes at level one, candidates at level two, and elections at level three, it was impossible for all computers tested to run that model. Furthermore, a ‘normal’ regression analysis with all variables interacted with the type of election is not included in the analysis, since it is hardly interpretable. For the sake of clarity, it has been chosen to compare each province with the national election (so only voters from a certain province) and the provincial election of that province.
24
These two variables are interacted with position on the list, gender, candidate ethnicity, first male and female on the list, age, age squared, experience, and party specific control variables.
20 If the question is whether people vote different in both elections, then a longitudinal study should be conducted with at some points in time interviews with the respondents.
21
4. Results
4.1 Descriptive information
The descriptive information about the Dutch general elections are included in Appendix B. For the Dutch parliamentary elections of 2012 there were in total 972 candidates competing for 150 seats. 7.6 million voters voted on the leading candidate of all parties that participated in the Dutch general election of 2012. In appendix B, the information about all leading candidates from all 21 political parties are presented. This information is in another way – and only for the eleven elected parties – presented in Figure 325.
Figure 3: Vote share leading candidates (municipalities, regions, provinces) of elected political parties (Dutch parliamentary elections of 2012)
For the eleven elected parties, on average 81,1% of the voters voted on the leading candidate. Of all elected parties, on average, the leading candidate from the Freedom Party (Geert Wilders) got the highest vote share (more than 93% of the votes for the PVV are for the number one, and at least 86% in each municipality of the PVV voters voted for Wilders). On the other hand, Sybrand van Haersma Buma got on average only 65% of the votes for his Christian Democratic party.
25
Other parties are excluded, since these parties get very few votes and in some municipalities even zero votes. There was one political party that only competed in one municipality (NXD in Amsterdam). It got only 62 votes, the lowest number of votes of all 21 parties.
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 L e a d in g ca n d id a te 's vo te sh a re o f p a rt y in l o ca lit ie s
VVD PvdA PVV CDA SP D66 GL CU SGP PvdD50PLUS
Municipality Region
22 Table 1 shows that apart from Van Haersma Buma all leading candidates of elected parties get from each municipality minimally 49% of the votes. Van Haersma Buma had only 18% of the vote share in one municipality in the Netherlands.
As the vote share for the leading candidates per region and per province is presented in the Figure 2, it does not take much difficulty to see that logically differences in shares between municipalities are larger compared to differences in shares between regions as well as differences in shares between provinces, as the latter two are weighted averages of municipalities. Regions are a good way to show that the 18% for Van Haersma Buma in one municipality in the Netherlands is not exceptional, since the party leader of the CDA gets in five regions (from the 40 regions in the Netherlands) less than half of the votes for his party. This means that in these regions, it could be possible that regional voting is high there.
Figure 4: Mean vote shares of candidates (number one through 49 on the ballot) of all elected parties per number on the list in their home municipalities versus the mean vote shares in all other municipalities (Dutch parliamentary elections of 2012)
The mean shares of all candidates from number one through number 49 on the list for all municipalities from elected parties are presented in Figure 4. It is divided in the share of the mean candidate’s home municipality share and the mean share of all other municipalities per position on the list. The line is the difference between the mean score in the home municipalities of the candidates versus the mean score of all other municipalities of the candidates. In Figure 4, the largest
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 Me a n vo te sh a re 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 Position candidate on ballot
Home municipality Other municipalities
23 difference is for the number two on the ballot and then it decreases for candidates lower on the list. However, the mean share in the home municipalities of the candidates is for all positions on the list larger compared to the mean share of all other municipalities in which the candidates do not live. Since most elected parties do not have a ballot with more than approximately 49 candidates, only the first 49 candidates on all lists are presented.
4.2 Global results
Table 1: General test of the locality and proximity hypothesis (Dutch general election) % votes candidate from party
(municipality) (ln)
Model 1 Model 2 Model 3 Model 4 Model 5
Candidate lives in municipality 9.574*** 6.194***
Candidate lives in region 5.141*** 2.192***
Candidate lives in province 2.675*** 1.680***
Distance in hectometres between municipality of living candidate and municipality of living voter (ln)
-1.319*** Constant -12.752*** -12.952*** -13.135*** -13.233*** -3.819*** Number of candidates 970 970 970 970 970 Number of localities 1 - 415 1 - 40 1 - 12 1 - 415 1 - 415 Number of observations 347637 33565 10100 347637 347637 Legend: * p<0.05; ** p<0.01; *** p<0.001
The first analysis consists of only the main variables of interest, with the candidate’s vote share in a municipality (logit) as dependent variable. The models, presented in Table 1, consist of all 970 candidate MPs for the Dutch general election. As the number of localities show, not all candidates are competing in all localities. There are candidates only competing in one municipality, or region, or province, as well as (most) candidates competing in all municipalities, regions, and provinces. As there are in total 972 candidates competing, the two candidates that are not included in this analysis, are the candidates only competing for votes in Bonaire, Sint Eustatius, and Saba. This means that people in the European part of the Kingdom of the Netherlands do not have the ability to vote on these two candidates and therefore are excluded from this analysis.
If the only variable in the estimation is whether the candidate lives in the municipality of the voter (yes or no) is included, it turns out that a candidate gets a significantly higher share in his municipality of living (listed on the ballot) compared to all other municipalities in which the candidate is competing. This is also true if only the region of living is estimated, and if only the province of living is estimated. Including these three variables in one model, it turns out that the lowest level of locality (municipality) has the highest estimated coefficient.
The other way to see whether there is an influence of local voting behaviour on a candidate’s vote share is to estimate the distance between the municipality of living of the candidate and the municipality of living of the voter. In Model 5 (Table 1), this is the only independent variable. It is highly significant with a negative number. This means that the further the distance between the
24 candidate’s place of living and the voter’s place of living, the lower the vote share of the candidate in the voter’s municipality of living. Hence, living in the municipality of the voter results in a significant higher vote share compared to living for instance 100 kilometres away from the voters in that specific municipality.
The first global results show that it is acceptable to study this subject in more detail. Therefore, in the next paragraphs, the results for each hypothesis will be shown.
4.3 Results hypotheses
The first hypothesis is the shared locality of living hypothesis. This was globally tested in Section 4.2. However, it is better to look at this hypothesis with other variables of interest and control variables. The results are presented in Table 2 in which Model 6 is the base model. This means that all other models with the question whether the candidate lives in the locality (yes or no) contain of all these variables. As there are a lot of control variables included in each estimation, only the variables of interest are included in this table. The full models are presented in Appendix C.
The results, now with control variables, are somewhat different compared to the global test from Section 4.2. If the control variables are added, then living in the region of the voter has the highest positive effect, living in the same municipality or province has lower (however also positive and significant) coefficients. This means that instead of the hypothesized largest effect of the lowest locality (municipality level), the highest positive effect is at the regional level. Living in a certain region is the key to success. However, living in a certain municipality (significant positive effect) should be added to living in a certain region and province.
Looking at the candidate and locality control variables, it is interesting to see that the effect of living alone in a certain municipality is not significantly different from living with one or two other candidates in the same municipality. From three or more other candidates living in the same municipality, the effect is negatively significant compared to whether a candidate is the only person from that party living in that municipality. The regional effect becomes negatively significant from two or more other candidates living in the same region onwards, showing that two or more candidates living in the same municipality is also negatively compared to living alone in a certain municipality, as the number of candidates in one municipality is never lower than the number of candidates in the corresponding region. At the provincial level, the effects compared to only one candidate living in a certain province are not significant.
In addition, a female candidate gets on average more votes, ceteris paribus, and six or more years experience as MP or minister gives a significant higher vote share, holding all other variables constant. Candidate MPs high on the list get more votes compared to candidate MPs lower on the list. Party control variables are not shown and are not individually significant. However, without these
25 variables, the prediction of votes (based on the model) for various real candidates was not good. Therefore, the fact that number one candidates get on average more than 80 percent of votes than other candidates is taken into account.
Table 2: Results hypotheses general election % votes candidate from party
(municipality) (ln) Model 6 (H1) Model 7 (H2) Model 8 (H3) Model 9 (H4) Model 10 (H5) Model 11 (H6)
Candidate and locality specific variables
Candidate lives in this municipality 1.782*** 1.685*** 2.484***
Candidate lives in this region 2.666*** 2.580*** 2.627***
Candidate lives in this province 1.753*** 1.831*** 1.757***
Distance between municipality of living candidate and municipality of living voter in hectometres (ln)
-0.967***
Candidate lives not in the election unit -0.279 -0.269 -0.268
Candidate was born in this municipality 0.804**
Candidate was born in this region 0.302**
Candidate was born in this province 0.337***
Candidate was born outside of the election unit
0.403
Distance between birthplace candidate and municipality of living voter in hectometres (ln)
-0.536***
Voter lives in the Randstad -0.289***
Voter lives in the Randstad * Candidate lives in this municipality
-1.398*
Distance between voter's municipality and The Hague in hectometres (ln)
-0.123***
Distance voter & The Hague * candidate lives in this municipality
0.665*** Constant 0.022 7.871* -0.135 4.628 -0.102 0.907 Number of candidates 150 150 150 150 150 150 Number of localities 415 415 415 415 415 415 Number of observations 62250 62250 62250 62250 62250 62250 Legend: * p<0.05; ** p<0.01; *** p<0.001
The second hypothesis, about geographical proximity, is estimated through one variable. It is expected that the larger the distance between candidate and voter, the lower the vote share of that candidate in the municipality of the voter. The coefficient is negative, which means that hypothesis 2 is accepted, since it turns out that a candidate should live near the voter to have a higher likelihood of votes in that municipality.
The third hypothesis is about the locality of birth of a candidate. The estimation results show that the effect is minor positive. This means that voters that live in the municipality of birth of a certain candidate will vote more on that candidate compared to the case that they live not in that municipality (controlled for locality of living of the candidate). Comparing this with the locality of living effects (which remain highly significant and change only 0.1 in coefficient), the locality of living is more important than the locality of birth. Nevertheless, the expectation was that the municipality of birth is the main level for a significant result. This expectation is true, but the effect is small.
26 The fourth hypothesis is that the larger the distance between the place of birth of the candidate and the locality of living of the voter, the lower the likelihood of a vote on that candidate. The estimated coefficient is highly negative, as expected. Therefore, hypothesis 4 is accepted.
The estimation results of the model for the fifth hypothesis about living in the Randstad versus the rest of the Netherlands, show that the regional effects are lower for candidates that live in the Randstad. Adding these variables results in a higher positive municipality of living-effect compared to the previous models. This model is an improvement compared to the base model (model 6) (Chi2=61.31 p<0.0001), hence it fits the data better.
The sixth hypothesis is that the distance between the voter’s municipality of living and the centre of politics (The Hague) is positively correlated, hence the further away a candidate lives from The Hague, the higher the expected positive effect on its vote share in a certain municipality. The results in table two show that the larger the distance between The Hague and the place of living of the candidate, the higher the likelihood of getting a higher vote share in his or her own municipality of living. This means that a candidate can better live in for instance Groningen, or Friesland, or Limburg, since the distance from these provinces to The Hague is larger than for instance living in the province of Utrecht.
The locality hypothesis are also tested for the twelve provincial elections, hence hypothesis 7. It is expected that the lower the level of a shared administrative community, the more positive the effect on the candidate’s vote share compared to all other localities. In this case, only the municipality of living as well as the region of living are included, since in most provinces almost all candidates live in the province. It turns out that the locality of living is a positive significant effect in all provinces. This effect is not only positive significant for the elected candidates in these provinces, it is even a higher positive coefficient when all candidates per province are included.
The collected information (birthplace, age and experience) about the candidates is in most provinces and cases not significant. Also in provinces with full data availability is this not the case. Therefore, this hypothesis for provincial elections could not accepted. This effect was significant, but small at the national level. The locality of living is in provincial states more important for a higher vote share than the place of birth (as this was also the case for the national election). Probably, because birthplace information of many candidates in many provinces is not available for most voters.
Although not hypothesized, it becomes interesting to make a first attempt to compare these two types of elections. As already discussed in Section 3.4 (Estimation restrictions), it is not possible to deal with the two structural differences between these two types of elections. Therefore, it is important to emphasize that the comparison is only a start for further necessary research.
27 Table 3: Example of estimation results provincial elections
% votes candidate from party (municipality) (ln) PS Friesland TK Friesland PS Gelderland TK Gelderland
Candidate lives in this municipality 6.799*** 3.781*** 6.345*** 6.417***
Candidate lives in this region 4.186 0.888* 2.368*** 4.055***
Candidate lives in this province (TK 2012) - 7.268*** - 0.663
Candidate lives not in the election unit -0.269 -0.102 0.453 -0.33
Constant -0.589 1.907 0.52 0.794
Number of observations 9531 20466 22018 42298
Number of candidates 353 758 408 761
Municipalities 27 27 18 - 56 18 - 56
Legend: * p<0.05; ** p<0.01; *** p<0.001
For every province, the first model estimated was with the elected candidates (excluding candidates with missing data). The second and third models are estimations without place of birth, age, and experience. The second model has the same sample as the first model and the third model consists of exactly the same variables as the second model, however now with all candidates competing in the provincial election.
The estimation results of model one and model two are in almost all provinces very similar. That shows that the inclusion of place of birth, age, and experience of candidates does not influence the estimated coefficients of the variables indication the municipality of the candidates . The variables of interest are in almost all cases and in almost all models highly significant. However, when the results of model 2 are compared with the results of the same model with a larger sample of candidates (model 3), it can be seen that the variables of interest have a much greater positive effect in model 3.
The results shown in table 3 are the estimation results of the variables of interest from model 326. This model is compared to a model for each province based on the national elections of 2012. Instead of using only the 150 elected MPs, to compare it with all candidates from the provincial elections, in the general elections also all the candidates competing in the province of interest are included.
28
5. Conclusion and discussion
This thesis researched the question under which circumstances people vote on a candidate from their own region. Many varieties have been taken into account. Yet, it turns out that regional voting behaviour matters – even in the least-likely case of the Netherlands. This country has no institutional features ensuring regional voting behaviour. It seems nevertheless that this way of descriptive representation is meaningful for many voters. A detailed estimation procedure made it achievable to see under which circumstances people do vote on a candidate from their own area. A candidate should live in the same province of the voter to get a significant higher chance on a vote from that voter – if the voter will vote on the party of the candidate. If a candidate is from the right party and living in the voter’s region, the likelihood of a vote will increase. Finally, a candidate gets the highest share in his or her municipality of living, as the provincial, regional, and municipal effects have all three a positive significant effect.
Different from other studies is that the shared administrative community is not the only estimation procedure to see to what extent there is regional voting behaviour. In addition, the proximity between a candidate’s place of living and the place of living of a voter has been estimated. It turns out that geographical proximity is also highly significant in the expected direction. This makes the results more robust. Comparing the results of this thesis, it is interesting to see that Van Holsteyn & Andeweg (2012) did not find an effect of geographical proximity in the national elections in the Netherlands in 2010.
It matters furthermore how many candidates from the same party are competing in the same area. If there are too much candidates living in the same area, they do not get a significant higher vote share compared to other areas anymore. For political parties it does therefore not electorally make sense to include too many candidates living in The Hague, Amsterdam, and other cities in the Randstad.
It has been found that there are regional differences. The further away a candidate lives from the centre of politics, the higher the vote share in his or her own municipality of living. For people not living in the Randstad, it could be more essential to vote on a person of their own region, as there are less people from these regions elected in preceding elections. In this thesis, it was found that people living not in the Randstad vote more on candidates living in their locality compared to people living in the Randstad.
Analysing the general elections with the provincial elections in the Netherlands, it turns out that the locality of living effect is the highest when all candidates are taken into account. For candidates it can be useful that they have room for a personal campaign in regions without other candidates.
