1Ségolène Royal, former socialist candidate at the presidential elections in 2007, was in for an unpleasant surprise when she stood again in the 2012 legislative elections. She was heavily defeated by a dissident socialist candidate (receiving 37.02% of the votes versus 62.98% for her opponent). This result was all the more surprising as a month before the socialist candidate for the presidential election, François Hollande, had received 55.59% of the votes in the same constituency. Conversely, although UMP candidate Nicolas Sarkozy received the majority of valid votes (50.35%) in the Boulogne Billancourt constituency in the first round of the 2012 presidential election, several weeks later his former Interior Minister and official UMP candidate, Claude Guéant, failed to get through to the second round in the same constituency by 137 votes. Beyond these two notable instances in which dissenting candidates won the vote, the discrepancy within the same constituency between the winning candidate at the legislative election and at the presidential election is far from marginal. For example, the socialist MP Jean-Louis Dumont was re-elected in a constituency where Sarkozy garnered 54.6% of the valid votes. In fact, in 49 constituencies (located in 28 départements) where Sarkozy gained more than 50% of the valid votes in the second round of the presidential elections, the winning candidate at the parliamentary elections was not a member of the broad center-right coalition. The parliamentary candidates elected in constituencies which leaned to the right at the presidential election represented the diversity of the political spectrum, from the communist party to the National Front. The converse is also true: MPs who did not belong to the presidential leftwing coalition were elected in constituencies which leaned to the left at the presidential election: 36 such candidates were elected in 27 départements, despite the fact that Hollande received more than 50% of the valid votes in these constituencies. The election of centre-right Jean-Christophe Lagarde is probably the most striking example in a constituency where Hollande won 66% of the vote. Overall, in 85 constituencies – i.e. 15% of French constituencies – there is a form of split ticket voting, [2] or at least a mismatch between the outcomes of the presidential and parliamentary elections. These examples cast doubts on two common features associated with French parliamentary elections in the literature: (i) They are usually considered as driven by the result of the last presidential election: these are the so-called “honeymoon parliamentary elections”. [3] According to this perspective, MPs are elected “riding on the coat-tails” of the new president. Fauvelle-Aymar, Lewis-Beck and Nadeau crudely sum up this idea: “‘presidentialisation’ has lessened the value of National Assembly elections, rendering them second order, increasingly a pawn in the presidential power struggle”; [4] (ii) Parliamentary elections are considered as driven by partisanship. Voters’ identification with a political party determines their choice of candidate, depending on the latter’s party affiliation. [5] In his longitudinal study of the determinants of MPs’ re-election from 1871 to 2002, Loonis (2006), states that “the political label has an increasing role after WWII […]. So it is logical to make the assumption that the role of individual features has declined.”

2In fact, previous examples illustrate that voters do not simply vote for a party or for/against the French president during French parliamentary elections. A candidate effect could be hypothesized to explain the discrepancy between presidential and parliamentary voting. This insight is far from being new. As early as 1966, studying the outcomes of presidential elections, Donald Stokes, one of the Michigan model’s advocates, underscored that the “fluctuations of electoral attitudes over these elections have to a remarkable degree focused on the candidates themselves”. [6] This candidate effect is defined as “the additional vote that a particular candidate wins (or loses)” [7] directly or indirectly because of his/her personal features. Since this seminal work, along with the growth of the “personalization of politics”, [8] the candidate effect has been increasingly studied, initially in relation to presidential elections in the United States. [9] Beyond this, with the rise of the concept of the “presidentialisation of politics”, the candidate effect has also been investigated in several Western countries in relation to the impact of the party leader’s image on the outcome of parliamentary elections. The conclusions of these studies have been disputed in terms of empirical results. [10] After a first wave of studies [11] that argued that a political leader’s image is relevant to voter choice and therefore has become more important, a second wave of studies challenged these results. [12]

3Curtice and Holmberg (2005) concluded that the independent effect of leader evaluation in six Western European countries was very limited and showed no increase over time. At the constituency level, the candidate effect has also been scrutinized focusing on the “personal vote”, [13] i.e. the portion of a candidate’s electoral support which stems from his/her own personal features, actions and record. [14] Although relatively scarce, studies have not been limited to majority electoral systems. [15] Nevertheless, the results did not unequivocally support the empirical relevance of the personal vote: some studies [16] clearly underscored it whereas others displayed inconclusive results. [17]

4French researchers have been reluctant to follow this trend in research [18] and have systematically analyzed the candidate effect. Nevertheless, scholars have shown that voters’ choices as well as the outcomes of presidential elections have been heavily influenced by a candidate’s image. [19] The impact of the candidate effect on voter choice and the electoral outcomes of parliamentary elections have been even more neglected. Although single member majority system elections with two rounds lead to strong hands-on campaigns in relatively small constituencies, existing studies characterize French parliamentary elections as a very unlikely case for the candidate effect: choosing an MP in France is described as heavily driven by partisanship, ideology, and the presidential elections. Is there any room left for a candidate effect? Analyzing electoral choice in French parliamentary elections is therefore a stringent test for the candidate effect hypothesis, which might bring new insights regarding the ongoing theoretical debate about the relative importance within electoral choice [20] of party identification (and ideology), party leader’s image (and/or presidential election effect), and the evaluation of candidates, as well as the relationship between these variables. [21] For this purpose, we specifically designed a survey for the 2012 French parliamentary elections to investigate the various determinants of electoral choices in parliamentary elections as well as the role of candidate evaluations.

5Beyond analyzing the candidate effect for the first time in French parliamentary elections via dedicated survey data, our paper also aims to disentangle the candidate effect from the main causal arguments. Section 1 offers a theoretical discussion of the potential benefits and pitfalls of integrating the candidate effect into explanations of electoral choices. Section 2 presents the survey on which this article is based and introduces our operationalization of dependent and independent variables. Section 3 displays the results of logistic and conditional logit regressions – which unequivocally support our expectation of a candidate effect – and illustrates the possibilities of this approach for assessing “the effects of a cause”.

Section 1 – Beyond endogeneity: a candidate effect in electoral choice?

6Even if the Michigan model included “personalities” as one of the short-term factors in electoral choice, Campbell et al. emphasized that voting behavior is first of all explained by “the role of enduring partisan commitments in shaping attitudes toward political object(s?)”. So-called party identification is a long-term affective orientation to a political party which is rooted in early socialization and based on an objective location in the social structure. [22] Party identification as an “unmoved mover” has “a powerful and pervasive influence on perceptions of political events”. [23] Some of the Michigan model followers underscored that “the specific effect of candidates’ image should thus be established after controlling for the other long-term factors (ideology, religion and class for example) and short-term factors (issues, particularly the economy) affecting electoral choice”. [24] The underlying idea is that as partisan identification filters and biases political perception, candidate evaluation among other factors is strongly determined by partisan identification.

7Nevertheless the relevance of party identification has been challenged on several grounds.

8(i) First, scholars have shown that there are more party switches than expected in response to short-term changes. [25] Thus the true meaning of what is measured by the classic variable of partisan proximity has been disputed. Furthermore, scholars point to the endogeneity problem. [26] Page and Jones underscored that “even short of major realignments, party affiliations are effects as well as causes in the electoral process”. [27] Sensitivity to certain issues – such as the economic situation – is usually thought to affect partisanship. But Page and Jones added that party loyalties “do not function purely as fixed determinants of the vote; those loyalties can themselves be affected by attitudes toward the current candidates”. In a recent study of Britain, Germany and the Netherlands, Garzia concluded that “when it comes to partisanship, voters’ evaluation of party leaders appears to have become the most powerful driver of partisan alignments at the individual level”. [28] Thus, according to him, including both partisanship and party leader evaluation as well as “the bidirectional arrows between these variables” in electoral research is necessary to adequately test the personalization hypothesis.

9(ii) Second, scholars questioned the relevance of partisan identification beyond the US. For multiparty systems or short-lived parties, some authors doubted the explanatory power of partisan identification [29] and stressed the importance of ideology instead of party. In Western Europe at least, the left-right dimension drives electoral choice. [30]

10So on a general theoretical level, electoral choice in parliamentary elections depends at a minimum on partisanship (which depends on the evaluation of party leaders), on the evaluation of party leaders (which depends on partisanship), and on ideology. It is also very likely that ideology, partisanship and the evaluation of party leaders are endogenous. Thus studying the personalization hypothesis at the level of party leaders is tricky. However, it might be studied more easily at the constituency candidate level. In nationalized political systems, if we accept that the evaluation of party leaders affects partisanship (as well as ideology) and vice versa, we might easily assume that voter evaluations of constituency party candidates do not affect partisanship (and ideology) even if partisanship (and ideology) is assumed to affect voter evaluations of constituency party candidates. Studying the personalization hypothesis is thus methodologically far easier at the constituency candidate level once a method for studying the candidate effect has been rigorously specified.

11In majoritarian Western parliamentary elections, the candidate effect is usually studied by the “personal vote” approach. In a nutshell, if the normal partisan vote refers to the electoral support a candidate at parliamentary elections can expect according to the underlying distribution of partisanship, of ideology, and of the support for the current government within her/his constituency, the “personal vote” is defined as the electoral support a candidate receives due to her/his own “performance” and qualities, independent of partisanship, of ideology, and of the support for the current government within her/his constituency. In more precise terms, Cain et al. defined the personal vote as, “[…] that portion of a candidate’s electoral support which originates in his or her personal qualities, qualifications, activities, and record”. [31] So, the personal vote is most often conceptualized as a “positive prejudice”. Whereas Blais et al. explicitly mentioned the negative side of the personal vote, they only operationalized the positive side by asking the following question: “Was there a candidate in your riding you particularly liked?”. Indeed, most scholars are looking for arguments that illustrate why incumbents [32] or contenders get more votes, but they rarely question the opposite situation, the case of a candidate generating votes against her/himself because of a negative perception of her/his person or actions. But the most severe limitation concerning this topic is a lack of precision. In 1974, Mayhew noted that since the end of World War II, there had been an increase in the proportion of votes received by congressional incumbents. [33] Since then, most of the time, the understanding of the personal vote has been connected (and limited) to the incumbency effect. [34] In one of their previous articles before their seminal publication on the personal vote, Cain et al. stated that their “analysis finds an incumbency advantage or personal vote in Britain which is much weaker than that in the United States”. [35] Although there is evidence of an incumbency effect, [36] studies of this phenomenon rarely explain exactly why incumbents get more votes but nevertheless consider it as a part of the personal vote. Moreover, not only incumbents but also contenders should be taken into account. If the former are better known in public, the latter may also gain some personal credit because of their rootedness in the local system; their participation in local networks; and/or, sometimes, simply their lack of prior involvement in politics…. Our study clearly aims to address the issue of eligibility defined sociologically: “in a given political configuration, eligibility is an addition of social and structural factors bringing and making it possible for a person to conquer – and to keep – a mandate within the framework of democratic competition”. [37] The various factors evoked in this definition concern three major kinds of largely interconnected phenomena: (i) Institutional aspects which affect access to office (the electoral system, local political context etc.); ii) Structural and contextual variables (geography and the size of the constituency, the candidate’s social background, and his/her relations with the constituency, its inhabitants and social institutions, and other common biographical variables [38]); (iii) Representations – of the candidates, the voters, the members of the political parties etc. – that are essential to understanding political phenomena in Western democracies. The last two dimensions are directly connected to the candidate effect but an inclusive definition of this particular effect is required to shed light on the eligibility issue.

12Thus the candidate effect is defined as “the additional vote that a particular candidate wins (or loses)” [39] directly or indirectly because of voters’ evaluations of him or her. The focus of this article will be on the direct (or local) candidate effect instead of the indirect (or national) party leader effect in legislative elections. Candidate evaluation refers to the features of a candidate perceived by voters. The perception of the candidate by the voter is usually called “image”. Several traits define a candidate’s image. Nevertheless, our definition of the candidate effect is limited to neither physical appearance [40] nor personality – narrowly defined by King as “their physical appearance, their native intelligence, their character or temperament, and their political style”. [41] Other than in France, traits and images of presidential candidates and party leaders have been investigated extensively. Nevertheless, there is still conflicting evidence, even in the US, about which traits are most salient for the evaluation of candidates and which are the relevant dimensions. According to Kinder, the four principal dimensions of presidential personality are competence, leadership, integrity, and empathy. [42] Miller et al. distinguished five “broad categories”: competence, integrity, reliability, charisma, and (idiosyncratic) personal quality. [43] Todorov et al. argued for three clusters based on seven traits: competence (competence, intelligence, and leadership), trust (trustworthiness, honesty), and likeability (charisma, likeability). Kessel’s typology of candidates’ qualities lists the following categories: general, record-incumbency, experience, management, intelligence, trust, and personality. [44] Shephard and Johns concluded tentatively that eight traits – competence, intelligence, leadership ability, charisma, likeability, attractiveness, honesty, and caring – have been used “most often in the literature and have been shown at least somewhere to correlate with electoral preference, however measured”. [45] Since there’s no general consensus on the most important traits, our study of candidate effect is based on voters’ evaluation of candidates on the eleven different traits we considered the most relevant (see below): general, likeability, concern, political renewal, political proximity, reliability, empathy, competence, leadership, record, local connection.

13These traits will allow us to test, all other things being equal, whether or not voters’ evaluations of candidates affect their electoral choice negatively or positively. Choosing a candidate in parliamentary elections will be understood as the product of partisanship, ideology, and the evaluation of both the party leader and the constituency candidate. However, we are less interested in (the relative weights of) the causes of voting behavior in parliamentary elections than in the effects of a cause – i.e. the evaluation of the candidate – in electoral choice in parliamentary elections. The highly polarized and presidentialized French political system is a demanding – and therefore perfect – test case for an empirical investigation of the candidate effect in parliamentary elections.

Section 2 – Design, data and measures

14Since one of the key objectives of the 1958 constitution was to limit parliament’s [46] prerogatives, the French political system – a semi-presidential regime – is usually described as highly presidentialized [47] in its various components: institutions, electoral competition, and party organization. [48] Consequently, parliamentary elections are seldom the primary focus for parties, voters, [49] and political scientists [50] in France. Nevertheless, the parliamentary component of the regime compels the president to appoint a prime minister who enjoys a majority in the lower chamber to ensure that the executive branch is unified and will be able to govern effectively. The goal is thus to avoid a cohabitation, which is a case of divided government stemming from the situation of split executive that occurs when the president and the majority in the National Assembly are rivals. [51]

15However, occurring soon after the presidential elections, parliamentary elections are usually considered to be just “elections of confirmation”. “In the French case, the institutionalization of electoral cycles that begin with a presidential election and are immediately followed by a honeymoon legislative election shifts the balance in favour of the president.” [52] The results of the presidential elections affect the parliamentary elections in that the presidential majority’s candidates benefit from them. Such a situation is greatly helped by the majority two-round system that produces a hugely disproportionate effect in favor of the winner. [53] Electoral choice in the “honeymoon parliamentary election” is driven by satisfaction with the new president. From a comparative point of view, the underlying logic of this French “honeymoon parliamentary election” does not differ fundamentally from the effect of voter evaluation of the party leaders in legislative elections in parliamentary regimes; the difference being that in France, the focus will only be on the evaluation of the president. The personalization hypothesis at the national level – the effect of voter evaluation of the party leaders – and the honeymoon hypothesis are de facto fused in the French case. The existing studies underscore the strength of their effect. As the (centralized) personalization effect is described as strong, [54] is there any room for a (decentralized) personalization effect, i.e. an effect of voters’ evaluation of the constituency candidates? Given the lack of investigation, scholars seem to consider this a worthless research agenda.

16The lack of research about the effect of voters’ evaluation of the constituency candidates probably also stems from another feature of the French political system. France is usually described as a polarized political system. In this perspective, electoral choice in French parliamentary elections is the product of this polarization. Nevertheless, there has been a fierce academic debate regarding partisanship versus ideology as the main cause of electoral choice in French parliamentary elections. Applying the Michigan model to France, Converse and Pierce strongly supported the idea that party identification is the main determinant of voting choice in French parliamentary elections. [55] But the explanatory power of party identification for the French case has also been questioned by some authors. [56] Instead, they have argued that ideology – the left-right dimension – is the determinant of electoral choice in France. [57] After an examination of legislative election surveys from 1958 to 1981, Lewis-Beck emphasized that “ideological identity appears generally to serve as the French voter’s compass”. [58]

17Finally, both determinants were acknowledged as relevant causes of electoral choice. [59]

18Scholars have underscored that electoral decisions in French parliamentary elections are driven by the voter’s evaluation of the president, by partisanship, and by ideology. Whilst these features do not prevent a candidate evaluation effect, they are unlikely to allow it either. Thus, if candidate evaluation does have an effect on electoral choice in one of the least likely cases [60] – the French parliamentary elections – then the candidate effect is likely to have a more general relevance as a cause of electoral choice. This is why we chose to study the 2012 parliamentary elections in France.

19These 2012 legislative elections took place in a specific context. In May 2012, Nicolas Sarkozy, the rightwing French president was defeated by François Hollande, his socialist opponent. This was only the second time in over 50 years that a socialist president had been elected in France under the Fifth Republic. [61] Nevertheless, despite a turnout of 80.4% of registered voters, the margin between the two candidates was closer than anticipated (Table 1).

Table 1

Results of the second round of the 2012 presidential election

Candidates Valid votes % of valid votes François Hollande 18,000,668 51.64 Nicolas Sarkozy 16,860,685 48.36

Results of the second round of the 2012 presidential election

20In such a context the importance of legislative elections was crucial. The turnout for the legislative elections dropped dramatically compared to the presidential election (about 30% less). The abstention rate reached 44.6 % in the second round which set a record under the Fifth Republic, confirming again that for French voters the election of the president remains the most important political appointment. And, once again, the presidential coalition won the majority of the seats at the parliamentary elections (Table 2).

Table 2

Results of the 2012 legislative elections compared to 2007

Party Seats 2012 Seats 2007 Communists 10 15 Other left 34 23 Green 17 4 Socialist 280 186 Centre + other right 39 35 UMP 194 313 Extreme right 3 1

Results of the 2012 legislative elections compared to 2007

Data

21The LEG 2012 survey [62] was developed in order to study the candidate effect during the French parliamentary elections in 2012. As in previous studies of parliamentary elections, [63] the survey is based on a sample of constituencies. Given the purpose of our research, the design of our sample departs from the previous ones in two key aspects. First, the number of constituencies has been drastically decreased in order to have a significant number of respondents per constituency. Second, we did not rely on random choice alone to select our constituency sample. A sample of 20 constituencies was selected on the basis of three criteria.

22(1) The aim was to distinguish between constituencies according to their main socio-economic features: their economic situation, whether they were rural or urban, their degree of industrialization. We ran a cluster analysis [64] of the metropolitan constituencies using the following variables as proxies: [65] the unemployment rate, the proportion of workers in agriculture, forestry and fishing, as well as the proportion of workers in industry. Five groups were distinguished with a similar weight. [66] Thus we selected four constituencies in each group.

23(2) The goal was to take into account the political leaning of the constituency. We distinguished leftwing and rightwing strongholds from constituencies in which the political competition between left and right was intense. We defined a leftwing (or a rightwing) stronghold as a constituency which had elected a leftwing (or a rightwing) incumbent at least four times out of five since 1988. Five constituencies were selected for each of the four different types of political feature.

24(3) Finally the constituencies had to reflect the geography of the metropolitan territory in order to take into account the various regional contexts and political cultures. The 20 constituencies that were chosen according to the first two criteria (Table 3) are located in 20 départements and 16 regions (Figure 1).

Table 3

Socio-economic and political features of the selected constituencies

competitive constituency rightwing stronghold with rightwing incumbent with leftwing incumbent leftwing stronghold %of metropolitan constituency urban, industrialized, average unemployement 1 1 1 1 20 rural, industrialized, low unemployement 1 1 1 1 23 rural, average industrialization, average unemployement 1 1 2 0 18 urban, weak industrialization, weak unemployement 1 1 1 1 19 urban, weak industrialization, high unemployement 1 1 0 2 20

Socio-economic and political features of the selected constituencies

Figure 1

Map of the French départements with selected constituencies

Map of the French départements with selected constituencies

25In each constituency, at least 150 phone interviews were undertaken by TNS–Sofres. The final database contains 3028 respondents. Given the number of interviews per constituency, simplified quotas have been used: gender, age (three categories) and occupation (three categories).

Measuring candidate evaluation: candidate traits and images

26As our focus is on parliamentary elections in a specific context, i.e. French politics, the eleven different traits studied do not simply replicate the traits usually included in the above-mentioned studies of candidates’ image: general, likeability, concern, political renewal, political proximity, reliability, empathy, competence, leadership, record, local connection. These eleven traits of course cover most of the various dimensions of candidate evaluation that have been studied in the literature. Nevertheless, as only the traits of competence, integrity and empathy have been studied so far in France, we had to create several specific questions in order to approximate the other traits (Table 4). The first six items are specifically linked to the constituency. Surprisingly, in a country in which the idea of national sovereignty has such resonance, voters perceive representation as primarily an electoral and territorial connection. [67] Further, MPs value local involvement more than parliamentary work. [68] The famous – and probably overestimated – feebleness of the French parliament, the electoral system, and/or the cumul des mandats are the usual explanations of these features. Thus, more than half of the items tackle traits linked to the constituency and not to the National Assembly. The national dimension of an MP’s role is captured, implicitly, by the other items. We also added several traits that are not usually included in candidate evaluation question batteries. First, an indicator of political proximity to the candidate has been used. A voter might be close(r) to a political party but might nonetheless consider that this party’s candidate is not close to him or her because of the internal heterogeneity of political parties. Local connection was introduced to take into account the perceived rootedness of a candidate. In French parliamentary elections at least, this quality is very often emphasised or denied. Finally, against the background of growing distrust of politicians, one item assesses whether or not the candidate embodies political renewal, i.e. a positive change in political personnel, whatever the reason (age, seniority, professional background, etc.).

Table 4

The traits of the candidates

Traits For each of the following, tell me whether the sentence characterizes (candidate’s name) very well, fairly well, rather badly, or not at all? General (s)he is made of the right stuff to be an MP. Likeability (s)he is nice. Concern (s)he worries me. Renewal (s)he embodies political renewal. Political proximity (s)he has ideas close to mine. Reliability (s)he really wants to defend our constituency. Empathy (s)he understands the problems of the people of our constituency well. Competence (s)he is able to fix the problems of our constituency. Leadership (s)he has the required influence to defend our constituency. Record (s)he has already proved her/his efficiency. Local connection (s)he is from here.

The traits of the candidates

27These eleven items were tested for a sample of party candidates. We selected 40 candidates that were the front-runners in the 20 selected constituencies as well as nine other candidates who were either the outgoing MPs or the third candidate at the end of the first round. Each respondent was asked about two or three candidates. Only the evaluations of UMP and PS candidates and their impact will be studied in this paper.

28As the sine qua non condition of any candidate effect is to know the candidate, the respondents were asked whether or not they were familiar with them without mentioning their party affiliation. In this regard, the UMP candidates were comparatively well known. Respectively, 87% and 79% of the UMP and PS candidates under study were known by the respondents. [69] Provided that the candidates were known, the eleven questions regarding the candidates’ traits were asked.

29Raw frequency results show that voters’ perceptions of candidates display strong variations. Tables 5 and 6 on the topic of local connection are a good illustration of these discrepancies.

Table 5

Results for local connection for PS candidates [70]

For the following sentence “(s)he is from here”, tell me whether the sentence characterizes (candidate’s name) very well, fairly well, rather badly, or not at all? Constituency number2 Very well Fairly well Rather badly Not at all Total 0702 27.48 35.11 15.27 22.14 100.00 1201 24.21 41.05 20.00 14.74 100.00 1701 15.17 34.48 17.93 32.41 100.00 1802 18.18 45.45 13.64 22.73 100.00 2105 18.39 37.93 20.69 22.99 100.00 2301 22.73 40.15 8.33 28.79 100.00 2501 33.93 37.50 10.71 17.86 100.00 2907 23.33 48.33 14.17 14.17 100.00 3302 19.23 48.46 15.38 16.92 100.00 5706 22.95 47.54 14.75 14.75 100.00 5919 30.23 36.05 13.95 19.77 100.00 6211 17.32 44.88 20.47 17.32 100.00 6604 24.11 41.96 15.18 18.75 100.00 9110 27.27 34.55 17.27 20.91 100.00 9209 15.79 38.16 23.68 22.37 100.00 Total 22.71 40.82 15.87 20.59 100.00

Results for local connection for PS candidates [70]

Table 6

Results for local connection for UMP candidates

For the following sentence “(s)he is from here”, tell me whether the sentence characterizes (candidate’s name) very well, fairly well, rather badly, or not at all? Constituency number Very well Fairly well Rather badly Not at all Total 0502 13.22 23.14 21.49 42.15 100.00 0702 21.62 23.42 18.92 36.04 100.00 1201 10.95 34.31 13.87 40.88 100.00 1309 23.20 40.00 13.60 23.20 100.00 2105 18.46 40.77 14.62 26.15 100.00 2301 13.53 22.56 15.79 48.12 100.00 2501 14.78 25.22 18.26 41.74 100.00 2907 13.83 28.72 18.09 39.36 100.00 3302 9.57 30.85 23.40 36.17 100.00 5706 10.29 33.82 23.53 32.35 100.00 6604 15.13 35.29 14.29 35.29 100.00 6801 22.73 44.70 14.39 18.18 100.00 6911 13.08 31.54 18.46 36.92 100.00 9209 11.72 26.90 21.38 40.00 100.00 Total 15.16 31.71 17.77 35.37 100.00

Results for local connection for UMP candidates

30The two tables illustrate a double tendency. (i) There is a difference in voters’ perceptions of candidates’ local connections as a function of whether candidates are from the left (PS) or from the right (UMP). Such an impression may be connected to the fact that the PS, having won the presidential elections, was the front runner for the legislative elections, (ii) but at the same time the noticeable internal variations of voters’ perceptions between candidates from the same party is a good indicator of an irreducible candidate effect (see for example the difference between candidates of the 2501 constituency and the 1701 constituency for the PS and between candidates of the 1309 constituency and the 2301 constituency for the UMP).

31As a multiple correspondence analysis displays only one dimension explaining 67% of the variance, a standardized additive scale has been built up with these 11 items. [71] For each individual, the Candidate evaluation variable contains the value associated with each candidate on the scale; the higher the value, the more negative the evaluation. For the first round, the evaluation scale for the socialist candidates is based on 16 different candidates (14 for UMP candidates). For the second round, the evaluation scale for the socialist candidates is based on 15 different candidates (13 for UMP candidates). As shown by Figures 2 and 3, socialist candidates are more positively evaluated than the UMP candidates.

Figure 2

Distribution of Candidate evaluation values for PS candidates

Distribution of Candidate evaluation values for PS candidates

Figure 3

Distribution of Candidate evaluation values for UMP candidates

Distribution of Candidate evaluation values for UMP candidates

32A second variable, Modified candidate evaluation, has been built by assigning the mean value (0) of the Candidate evaluation scale to the respondents who did not know the candidate or who were unable to evaluate the candidate. This variable will allow us test the robustness of the candidate evaluation variable on the whole set of respondents. [72]

Section 3 – Candidate effect under empirical tests

40We will use logistic regressions [78] to predict the vote for a party candidate (either UMP or PS) in the first or the second round according to the respondent’s evaluation of the party candidate, the respondent’s party ID, the distance from the party, and approval of the president’s performance.

41Tables 7 and 8 respectively present the results of these analyses for UMP and PS candidates for the first and second round. The reported odds ratios indicate the effect of the independent variable on the dependent variable.

Table 7

Effect of candidate evaluation on voting behavior for UMP candidates (logistic regression) [79]^*,**,***

UMP-1st round UMP-2nd round Respondents who evaluated the candidate All respondents Respondents who evaluated the candidate All respondents Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7 Model 8 Model 9 Model 10 Party ID 6.5*** 6.9*** 7.8*** 8.0*** 3.6*** 4.4*** 4.8*** 5.3*** Distance .82** .74*** .74*** .71*** .7*** .67*** .69*** .67*** Presidential approval 2.0** 2.7*** 2.4*** 3.0*** 4.2*** 4.7*** 4.3*** 4.8*** Candidate evaluation .05*** .08*** .05*** .07*** Modified candidate evaluation .07*** .07*** Pseudo R2 .40 .55 .39 .54 .43 .42 .61 .45 .57 .46 Log likelihood -367 -258.9 -349 -419.7 -514.3 -345 -219.6 -309.6 -341.2 -429.5 Percentage correctly predicted 82,8 86.8 83.9 88.6 86.9 82 89.7 83.9 88.5 84.7 N 951 907 907 1577 1577 870 824 824 1186 1186

Effect of candidate evaluation on voting behavior for UMP candidates (logistic regression) [79]^*,**,***

Odds-ratios are presented;
* p<.05;
** p<.01;
*** p<.001

Table 8

Effect of candidate evaluation on voting behavior for PS candidates (logistic regression) [80]^*,**,***

PS-1st round PS-2nd round Respondents who evaluated the candidate All respondents Respondents who evaluated the candidate All respondents Model 11 Model 12 Model 13 Model 14 Model 15 Model 16 Model 17 Model 18 Model 19 Model 20 Party ID 12.9*** 13.1*** 9.2*** 10.4*** 3.3*** 5.3*** 4.9*** 6.4*** Distance .90 .86** .87** .85*** .90 .84** .86** .84*** Presidential approval .62 .19*** .20*** .13*** .16*** .07*** .08*** .06*** Candidate evaluation .08*** .12*** .03*** .05*** Modified candidate evaluation .15*** .06*** Pseudo R2 .31 .49 .39 .45 .39 .47 .58 .40 .53 .43 Log likelihood -439.1 -315.0 -375.8 -542.1 -600.1 -314.5 -238 -338.1 -443.1 -536.2 Percentage correctly predicted 75.5 85 83.2 84.1 82.9 85.3 89.5 84.4 86 82.8 N 919 886 886 1464 1464 921 885 885 1374 1374

Effect of candidate evaluation on voting behavior for PS candidates (logistic regression) [80]^*,**,***

Odds-ratios are presented;
* p<.05;
** p<.01;
*** p<.001

42First we tested the effect of Candidate evaluation in bivariate logistic regressions in Models 1, 6, 11 and 16. In the four models, this independent variable is highly significant and in the expected direction. A standard deviation increase in the Candidate evaluation decreases the odds of voting for a candidate between 96% (for a socialist candidate in the second round) and 90% (for a socialist candidate in the first round). It also has strong explanatory power. Between 75.5 and 85.5% of the cases are correctly predicted and the PseudoR2 has a value between 0.31 and 0.47.

43Second we tested the impact of Candidate evaluation with the three classical variables in multivariate logistic regressions in Models 2, 7, 12 and 17. The models have strong explanatory power: between 85 and 90% of the cases are correctly predicted and the PseudoR2 has a value between 0.49 and 0.61. From this perspective, the multivariate models perform systematically better than the univariate ones. Whatever the party and the round, the odds ratios associated with the respondent’s party ID, are statistically significant and in the expected direction. When respondents have the same party ID as the party candidate, the ratio of the probability of voting for the party candidate over the probability of not voting for the party candidate is 6.5 and 12.9 higher respectively for UMP and socialist candidates in the first round than among the respondents without the same party ID as the party candidate. The odds ratios (respectively 3.6 and 3.3) are less powerful in the second round for both party candidates. This result is consistent with the severely reduced political choice in the second round of the elections. The distance variable is always in the expected direction but is never statistically significant for the PS candidates, whatever the round. When the distance between the respondent and the party increases by 1, the ratio of the likelihood of voting for the candidate over the likelihood of not voting for the candidate decreases by 18% and 30% for the UMP candidates respectively in the first and second round. Finally, Presidential approval is also always in the expected direction even if it is not significant in the model (Model 12) predicting the vote for a socialist candidate in the first round. The effect is particularly robust for UMP candidates: if respondents disapprove of the president’s performance, the odds of voting for a UMP candidate doubles in the first round and quadruples in the second round. Conversely, disapproving of the president’s performance divides the odds of voting for a socialist candidate by around 6 in the second round. The evidence also strongly supports the candidate effect hypothesis. The odds ratios associated with Candidate evaluation are highly significant and in the expected direction for both parties and both rounds. The effect is minimal for socialist candidates in the first round and maximal for the socialist candidates in the second round: when the candidate evaluation is worse by a standard deviation, the odds of voting for a socialist candidate is divided by 6 in the first round and by 20 in the second round. The average marginal effect [81] of the variables included in Models 2, 7, 12 and 17 confirms the strong impact of candidate evaluation compared to the other variables (Table 9). Whereas being socialist increases the likelihood of voting for a socialist candidate by 28% on average – the strongest party ID effect – a 1 point deterioration in the candidate evaluation scale decreases the likelihood of voting for a socialist candidate by an average of 23%. The effect of this variable is stable for socialist and UMP candidates, and in both rounds. Furthermore, in all models, the average marginal effect of Candidate evaluation does not vary much in relation to whether or not the respondent has the same party ID as the candidate. [82] All other things being equal, the effect of candidate evaluation on voting behavior is far from marginal.

Table 9

Average marginal effects (Models 2, 7, 12 and 17)

PS 1st round PS 2nd round UMP 1st round UMP 2nd round Party ID .28 .09 .16 .10 Distance -.01 -.01 -.02 -.03 Presidential approval -.05 -.15 .06 .12 Candidate evaluation -.23 -.24 -.22 -.21 Candidate evaluation when Party ID =0 -.30 -.31 -.26 -.25 Candidate evaluation when Party ID =1 -.29 -.23 -.30 -.25 N 886 885 907 824

Average marginal effects (Models 2, 7, 12 and 17)

44A way to assess the value added by candidate evaluation to explain voting behavior is to replicate the models without the candidate evaluation variable (Models 3, 8, 13 and 18) on the same sample of respondents. The PseudoR2 is systematically lower (between 0.1 and 0.19 less) as well as the proportion of correctly predicted cases (around 2% and 5% less respectively for the first round and the second round). The explanatory power of these models is only slightly higher than the univariate models (Models 1, 6, 11 and 16) testing the candidate effect hypothesis.

45As the previous models were tested only on respondents who evaluated the candidates, the robustness of our hypothesis should also be tested on the whole sample of voters. Using the Modified candidate evaluation, we replicated the previous models including the comprehensive set of variables as well as including only the classical variables. In Models 4, 9, 14 and 19, the odds ratio associated with Modified candidate evaluation is systematically highly significant and in the expected direction. [83] This means that a positive or negative evaluation of the candidates significantly affects the likelihood of voting for them, even in comparison with lack of knowledge of a candidate. In other words, electorally it is really better to be well known than unknown, but also better to be unknown than despised. The odds ratio associated with the classical variables are also in the expected direction and statistically significant. In Models 5, 10, 15 and 20, we replicate the analyses without Modified candidate evaluation. Whatever the party and the round, the odds ratios associated with the respondent’s party ID, the distance from the party, and approval of the president’s performance are statistically significant and in the expected direction. Their values do not significantly differ from those from Models 3, 8, 13 and 18. This is a valuable insight that means that the samples of respondents able to evaluate the candidates are not outliers and our results are robust. Again, this is also a way to assess the value added by candidate evaluation to explain voting behavior. The PseudoR2 is again systematically lower (between 0.07 and 0.11 less) as well as the proportion of correctly predicted cases (between 1.2% and 3.8%).

46Finally, we tested the robustness of our results by running models with fixed effects, including dummy variables first for the clusters [84] and then for the constituencies themselves. The variables testing the candidate effect are always significant and in the expected direction. [85] Hence our results do not depend on the specificities of any individual constituency or any cluster of constituencies.

47Thus evidence systematically and strongly supports the candidate effect theory. Variables testing it are always statistically significant and in the expected direction whatever the sample, the candidate, the round and the other variables included.