PARTY, IDEOLOGY, AND VOTE INTENTIONS: DYNAMICS FROM THE 2002 FRENCH ELECTORAL PANEL*

Transcription

1 PARTY, IDEOLOGY, AND VOTE INTENTIONS: DYNAMICS FROM THE 2002 FRENCH ELECTORAL PANEL* Éric Bélanger Department of Political Science McGill University 855 Sherbrooke Street West Montreal, Canada H3A 2T7 Tel.: ; Fax: Michael S. Lewis-Beck Department of Political Science University of Iowa Iowa City, Iowa Jean Chiche CEVIPOF Centre de recherches politiques de Sciences Po Institut d Études Politiques de Paris 98, rue de l Université Paris, France chiche@msh-paris.fr Vincent Tiberj CEVIPOF Centre de recherches politiques de Sciences Po Institut d Études Politiques de Paris 98, rue de l Université Paris, France vincent.tiberj@sciences-po.fr * Presented at Workshop No.22 : «Beyond Party Identification and Beyond,» ECPR Joint Sessions of Workshop, Nicosia, April 25-30, Do not quote or cite without permission.

2 PARTY, IDEOLOGY, AND VOTE INTENTIONS: DYNAMICS FROM THE 2002 FRENCH ELECTORAL PANEL Abstract The debate over the relative importance of ideology versus party for vote choice in France is enduring. Resolution of the debate would have much value, for the light shed on sources of stability and change in multiparty electoral systems generally. The main reason the debate continues is that previous studies examining that question have been plagued by difficulties pertaining to variable measurement, model specification, election type, and research design. We address these problems and provide new evidence from the 2002 French Electoral Panel. Most notably, these data allow stronger causal inference because party identification and ideological identification are both measured in the first wave of the survey, that is, before the declaration of vote actually occurs. We estimate a multi-equation model of first-round legislative vote intention as measured in the second wave of the panel using two-stage least squares, ordered logit, as well as binomial and multinomial logit techniques. The results indicate that ideological identification systematically outweighs party identification in shaping the French voter s choice.

3 French voters, like their American counterparts, have been the object of serious national election studies for some time. As in the United States, much of that research has sought to define the long-term social-psychological anchor that guides the French electorate. In the United States, the conclusion, accepted with minimal controversy, is that party identification serves as that anchor. Such a conclusion is more controversial for Western European democracies, where the nature and importance of party identification has been in debate for some time (see the following classic set of readings: Budge et al., 1976; Dalton et al., 1984; Franklin et al., 1992). Perhaps it is in France that the debate is most intense, with scholarly rivalry over the role of party identification versus ideological identification ongoing. Looking at the larger scheme of things, why is this issue important? Because it informs us about democratic system performance, in particular about the determinants of stability and change in multiparty polities. Imagine two opposite theoretical possibilities, for mass publics in any democracy. First, suppose party identification rules. Then we might expect to observe a system tending to stable, enduring parties. Second, suppose ideological identification rules. Then, we might expect to observe an unstable system, with parties coming and going as left (or right) electors seek to match their preferences. Of course, a third possibility is that a system is some mix of the two identifications. In that case, it is especially worth figuring out the relative weights of the two anchors, for it helps diagnose otherwise peculiar patterns. For example, it could help explain the characteristic weave of continuities and discontinuities marking a party system. It is worthwhile keeping in mind these larger questions, as we turn to the specifics of the debate in the French literature. On the one hand, it is argued that party identification is the dominant long-term force acting on the French voter. Converse and Pierce (1986) established this perspective, in their

4 2 application of a Michigan-style voter paradigm to the 1967 French National Election Study. Summarizing a large part of their analysis, they conclude as follows: Most of the time, in the most compelling tests, it appears as though partisanship is a more efficacious frame of orientation than is the sense of one s own left-right position. What we have never encountered to date is an instance in which left-right orientations clearly and persistently outweigh partisan attachments (Converse and Pierce, 1986, pp ). The question was reopened in Pierce s (1995, p.138) examination of the presidential election of 1988, where he basically comes to the same answer: the bulk of the evidence suggests that on the individual level, partisanship has a more direct impact on electoral choice than ideology. On the other hand, it has been argued that left-right ideological identification dominates. This perspective is widely held by French electoral researchers. Percheron (1977) made the case that party identification was irrelevant, and that what was relevant was left-right ideology, passed from parent to child. Haegel (1990, p.153) flatly claimed that the notion of party identification was not très prégnant in France, and that what matters is identification with the left or the right. Mayer (1997, p.20), in a consideration of recent elections, makes reference to the fact that le clivage gauche-droite reste le principe majeur de structuration et de cohérence de l espace politique. Some U.S. researchers of the French electorate have made similar arguments. Fleury and Lewis-Beck (1993, p.1107) comment: Ideology, not party, is the premier psychological anchor of the French voter, according to this analysis of the 1967 French National Election Study. Work on subsequent elections ( ) by Lewis- Beck and his co-authors reinforces this finding (Lewis-Beck and Skalaban, 1992; Lewis-Beck and Mitchell, 1993).

5 3 Recent work, using more current data and hitherto untried estimation techniques, has sought a way out of this impasse. Lewis-Beck and Chlarson (2002, p.491), examining 1995 presidential election survey data (N = 4,078), explicitly aimed to reconcile the opposing perspectives on the role of party and ideology. They find, in a comparative logistic regression analysis, that party dominates on the first ballot while ideology dominates on the second ballot, concluding French voters appear truly bipolar in their long-term attachments. (Lewis-Beck and Chlarson, 2002, p.511). However, Evans (2004, p.68), in his maximum likelihood investigation of this question using the 1993 legislative first ballot choice, finds that [p]arty identification, where it exists, is a stronger voting anchor than ideology, and is the principal cue provider. In sum, the question is party or ideology the more important anchor for the French voter? remains. We believe the issue persists for several reasons, among them variable measurement, model specification, election type, and research design. There have been considerable difficulties with the measurement of party identification, in particular. The issue of model specification concerns not only what independent variables should go into the model, but whether the causal links are reciprocal. With respect to election type, there has been no consistency, as sometimes presidential, sometimes legislative elections are examined, on the first or the second ballot. The research design heretofore, but with one exception, has been static, cross-sectional. This is of special concern given the causal issues involved. Below, we address these problems, then estimate a new model on a unique dataset, the 2002 French Electoral Panel. The results, we believe, go a long way toward laying this debate to rest. Measurement Issues

6 4 In French election surveys, party identification and ideological identification are regularly measured. The trouble is, at least with the first, the measurement varies from study to study. There are two basic ways of posing these identifications to respondents: closedended or open-ended. Ideology is virtually always offered as a closed-ended seven-point selfplacement item, of the following style: Going from left to right, where would you personally place yourself on this scale? [The interviewer shows the respondent a seven-point scale running from left to right.] Such a question has been regularly employed in French surveys over many years, with considerable success. Invariably, 90 percent or more of respondents can so place themselves (Lewis-Beck and Chlarson, 2002, p.494). For example, in the 1995 French National Election Study, 98.1 percent of those interviewed selected one of the points. Some analysts exclude those in the center position of 4, and some do not. The argument for exclusion is that it is not a meaningful choice, reflecting instead low political interest. The argument for inclusion is that it faithfully represents those who are ideologically moderate. Evans (2004, p.56), in a current treatment of this question, concludes that we do not feel that to include the centre in an ideological measure is any more contentious than excluding it. Turning to party identification, it may be asked closed-ended, e.g.: Here is a list of political parties. Could you tell me which you feel closest to? [The interviewer shows the respondent a list of parties.] But it may also be asked open-ended, e.g.: What is the party you feel closest to? [The interviewer writes down whatever the respondent says.]

7 5 The difference in format, it turns out, is crucial (Evans, 2004; Lewis-Beck, 1996, p.517). With the closed-ended item, party identification estimates can easily run in the 70s. With the open-ended item, the estimates tend to not go beyond the 50s. For example, in the 1995 French National Election Study, an open-ended question was used, producing a party identification estimate of 46.9 percent (Lewis-Beck and Chlarson, 2002, p.493). In contrast, the closed-ended item in the same survey yielded an estimate of 86.6 percent (Evans, 2004, p.50). These large differences occur because the closed-ended item is essentially viewed as a vote intention question by respondents. Indeed, such a closed-ended party identification item has been shown to correlate.94 with vote intention (Lewis-Beck, 1996, p.522). These items on party and vote are tautological or, put another way, closed-ended party identification is highly endogenous, a causal product of vote preference. In sum, measures of left-right identification and party identification are routinely available in French election surveys. However, if closed-ended items are to be used, issues of endogeneity need to be addressed, especially for party identification. Model Specification To assess correctly the relative importance of party and ideology, model specification should tend to two things. First, the vote equation should contain other relevant independent variables, in order to reduce any estimation bias in these relative effects. And, second, it should take into account any reciprocal linkage among these independent variables, namely party identification and ideological identification. With respect to the first point, it would seem the equation should have the following general specification: Vote = ƒ (Party, Ideology, Cleavages, Issues) Eq. 1

8 6 Social cleavages ( les variables lourdes ), such as religion and class, represent other perennially important long-term forces acting on the French voter (Boy and Mayer, 1993, p.174). Issues represent short-term forces, particularly relevant to the election under study. For the 2002 elections, issues of economic well-being, law and order, and immigration appear to have been at the forefront of voter s minds (Mayer and Tiberj, 2004, p.36). With respect to the second point, it has been argued that there is a reciprocal link between party identification and ideological identification, one influencing the other (Fleury and Lewis-Beck, 1993; Evans, 2004). In fact, Evans (2004, p.54) contends this issue constitutes the major debate in the French electoral literature over the past decade. If such is the case, a multi-equation system is implied, and it is no longer possible to estimate Eq. 1 with ordinary least squares (OLS), since both the party and the ideology variables, in raw form, will be endogenous. Facing this difficulty, Fleury and Lewis-Beck (1993) employed an instrumental variables approach in a two-stage least squares (2SLS) estimation. By way of contrast, Evans (2004, pp.58-59) employed a full-information maximum likelihood technique to solve the problem. Regardless of which approach is followed, it seems clear that this threat of bias has to be overcome, in order to answer our central question. Election Type What elections offer the clearer test, presidential or legislative? First round or second round? Lewis-Beck and Chlarson (2002) examine the question using the 1995 presidential election, first and second ballots. But Evans (2004, p.57) argues persuasively that legislative elections afford a better test, because candidate personality is not so intrusive. In fact, the bulk of the relevant studies have looked at the legislative arena (Converse and Pierce, 1986;

9 7 Fleury and Lewis-Beck, 1993; Evans, 2004). But not all of them have looked at the same round of balloting. For example, Fleury and Lewis-Beck (1993) examine the first ballot, and a combination first-and-second ballot called the decisive vote, after Goguel (1968). Evans (2004, p.57) employs the first-ballot vote (recalled). If these studies are any guide, legislative contests should afford a better test. Further, the first ballot should offer a tougher test of the hypothesis that ideology dominates party. That is because on the first ballot, choice is likely to be sincere rather than strategic, with voters having more opportunity to pick their own party, since none has yet been eliminated (Elgie, 1996, p.61). Research Design As Evans (2004, p.54) claims, the great debate in the French electoral literature is over the reciprocal causality between party identification and ideological identification. When questions of causality are paramount, we seek research designs that allow the establishment of strong inference. Unfortunately, all available French election datasets have been cross-sectional, single slices in time. Thus, even though simultaneous equation systems with reciprocal links may be forced on them, as in the case of Fleury and Lewis-Beck (1993), and Evans (2004), the causal inference they offer remains weak. Their static design does not meet a critical condition of causality, i.e., X occurs before Y in real time. To meet this temporal condition, a dynamic design is needed, where the key variables are measured at different time points. In other words, panel data are what is lacking. Of course, Converse and Pierce (1986, pp , 328, ) did offer analysis of a French National Election Study panel. However, because of issues of small N and panel attrition, they decided to depend

10 8 primarily upon the 1967 sample for most of our estimates of the abiding characteristics of the French electorate (Converse and Pierce, 1986, p.802). In any case, their panel data, now over 35 years old, were never archived for use by other scholars. Happily, a contemporary panel has been conducted, covering the 2002 French elections, and has come available (for an English-language description of the characteristics of the panel, and its archiving, see Cautrès, Mayer, and Salomon, 2003). The panel was carried out in three waves: Wave 1, April 8-20 (before the first ballot of the presidential election), face-to-face interviews, N = 4,107; Wave 2, May (after the second ballot of the presidential election), telephone interviews, N = 4,017 (of whom 1,822 were interviewed in Wave 1); Wave 3, June (after the second ballot of the legislative election), telephone interviews, N = 2,013 (of whom 1,417 were interviewed in Waves 1 and 2). We use these 2002 French Electoral Panel data to assess which factor, party or ideology, matters more to the French voter s decision. Measures and Correlations The variables of most obvious interest are party identification, ideological identification, and vote. Items on these variables are not always repeated across the three waves. However, there are items sufficient for our critical test. In particular, the 1,822 respondents interviewed in both Waves 1 and 2, were asked party identification and ideological identification in Wave 1 and first-round legislative vote intention in Wave 2. [1] These core variables, in conjunction with a collection of exogenous and control variables from Wave 1, allow us to make the strong causal inferences that panel data afford (Finkel, 1995). Note the novel feature is that party identification and ideological identification are no longer

11 9 measured at the same time as vote intention. Instead, they are measured before the declaration of vote actually occurs. Incorporation of this temporal dimension should go a long way in sorting out the true relative importance of these variables. Before proceeding to model estimation, it is useful to detail the measures, and their simple associations. The dependent variable, legislative vote intention in the first round, is measured as follows: With respect to the legislative elections that will take place in the month of June, for which party are you most likely to vote on the first ballot? [The interviewer then reads the following list: Extreme Left (LO, LCR, PT); Communist Party; Socialist Party; Movement of Citizens; Greens; Other Ecologists; Hunters; UDF; RPR; RPF; Liberal Democrat; National Front; MNR; none; other.] Instrumentation for ideological identification follows the standard, seven-point, closedended item: One regularly places the French on a left-right scale. Reflecting on the political ideas that you have held since age 18, how would you place yourself now on this scale? [The interviewer shows the respondent a seven-point left-right scale.] The question of the substantive meaning of a particular placement on this numerical scale has been extensively considered in earlier investigations. Lewis-Beck and Chlarson (2002, pp ) review this controversy, and demonstrate that, using the 1995 French National Election survey, left-right self-placement correlates highly (R =.50) with views on a battery of national issues. This result is essentially replicated with these 2002 data. Attitudes on the following policy items cutting taxes, privatization, the death penalty, making money, law and order, maintaining traditions, anti-globalization, poverty, social justice, equality of

12 10 chances, pro-statism again correlated highly (R =.54) with left-right self-placement. Thus, it appears that, substantively, left-right ideological placement serves as a summary opinion on a whole host of issues. Party identification is measured in a closed-ended question, with respondents selecting from a list. The item reads as follows: Here is a list of parties or political movements. Can you tell me which you feel closest to or, let s say, the least far from? [The interviewer shows the respondent the exact same list of parties read from in the vote intention question above.] As can be seen, the first two items, on vote intention and left-right ideology, are phrased in ways that have become the French standard. With respect to the party identification variable, however, there are two traditions of measurement, and the closedended format is the one available here. The expectation is that it would give an inflated estimate of party identification, and it appears to. Fully 78.7 percent of the respondents (Wave 1) selected a party from the above list, when asked to do so. Given what we know, this is clearly much more than an open-ended item estimate would produce. Besides this biased frequency distribution, we would expect that the correlation between party identification and vote intention would be biased upward. Coding partisan responses for a complex multiparty system such as the French one is never easy. However, there are some rather accepted combinations which simplify the task (this coding is discussed further in the footnote). [2] The partisan responses for these two variables we code as follows: Extreme Left = 1, Communist Party = 2, Socialists and Movement of Citizens = 3; Greens, Ecologists, and Hunters = 4; UDF = 5; RPR, RPF, DL = 6; FN and MNR = 7. The correlation between these two variables, party identification (Wave 1)

13 11 and vote intention (Wave 2), is in fact large, r =.75. Indeed, it is somewhat larger than the comparable correlation for ideological identification (Wave 1) with vote intention (Wave 2), where r =.65. At first blush, it may appear that party influence exceeds ideology, from a comparison of these correlations,.75 >.65. However, this comparison is invalid because, even though both party and ideology are measured prior in time to vote, endogeneity problems remain. That is, vote likely influences party, while party influences vote. And, party may influence ideology, and vice versa. These reciprocal links pose a causal tangle. Furthermore, since the data are gathered at two points in time, the expectation is further bias from a correlated error problem, with factors influencing party (or ideology) at time t-1 also influencing vote at time t. The solution is to exogenize the party and ideology variables. Once they are effectively exogenous, then the simultaneous equation bias and the correlated error bias are eliminated, and least squares (or some comparable technique) can be properly applied to estimation of a vote model. Fortunately, a large number of exogenous variables are available in Wave 1, namely age, gender, schooling, income, occupation, work sector, organizational membership, democratic satisfaction, interest in politics, support for the European Union, 1995 Chirac vote, and 1995 Jospin vote. [3] Using them, we apply an instrumental variables approach, in order to arrive at usable exogenous proxies for party (P t-1 ) and ideology (I t-1 ). These variables are each regressed on a set of exogenous variables, in order to create predicted variables P* t-1 and I* t-1, which become effectively exogenous themselves. [4] Further, these variables appear to be good instruments, in fact equally good, with P t-1 and P* t-1 correlating at.60, and I t-1 and I* t-1 also correlating at.60.

14 12 By analyzing the relationship of P* t-1 and I* t-1 instead of P t-1 and I t-1 to vote intention, we can get closer to the truth of things. First, look at the simple correlations. When P* t-1 is correlated with vote, r =.54; when I* t-1 is correlated with vote, r =.56. The suggestion is that ideology is at least competitive with party in its influence. However, this is a very preliminary suggestion, since proper controls, in a multivariate context, have not been applied. It is to this task we turn. Model Estimation The general vote equation we wish to estimate is dynamic, incorporating the panel time lag, as follows: Legislative Vote Intention t = ƒ (Party t-1, Ideology t-1, Cleavages, Issues) Eq. 2 We measure legislative vote intention as described above. Party t-1 and Ideology t-1 we measure with the instrumental variables just developed in the above first-stage regressions, P* t-1 and I* t-1. With respect to cleavages, we specify the variables of religion and class, measuring the first in terms of the observance of religious services (4 to 0, from practicing believers to nonpracticing non-believers), and the second in terms of subjective social class (1 = middle, 0 = working). With respect to issues, we specify the economy (1 = unemployment is an important problem, 0 otherwise), immigration (4 = completely agree there is too much to 1 = completely disagree), and insecurity (1 = feel completely secure to 4 = not at all secure). For further details on the coding of these and the other variables, see the bottom of Table 1. The least squares estimates for this second stage (2SLS) equation, Eq. 2, appear in column 1, Table 1. Observe that all the coefficients are in the expected direction, and easily significant. Middle class and more religious people favor parties on the right, as do those who

15 13 worry about immigrants and insecurity. Concern over the economy, namely unemployment, pushes voters to the left. These findings are not surprising. Rather, they fit well within the expected pattern of results, especially for 2002 (see Lewis-Beck, 2004). Overall, the model accounts for the variance in vote preferences in that election rather well, with an adjusted R- squared =.43. We are reassured that there is strong support for this general specification. (Table 1 about here) What about the key question at hand? Upon examination of the regression coefficients, it is clear that ideological effects dominate party effects, respectively,.63 >.42. (This immediate comparison of the coefficients is possible because the metrics for the two variables are the same, each seven-point scales.) The same conclusion is arrived at when the standardized coefficients, the beta weights, are compared, respectively,.35 >.23. By these tests, ideology dominates. (Further, this conclusion sustains itself against challenges to the sample size; see footnote for details). [5] Of course, these estimates are only direct effects. There are indirect effects, via the influence of party on ideology, and vice versa. When these are taken into account, the picture might change. To estimate this possibility, it is necessary to move explicitly to a threeequation model, where: Ideology t-1 = ƒ (Party* t-1, Cleavages, Issues) Eq. 3 Party t-1 = ƒ (Ideology* t-1, Cleavages, Issues) Eq. 4 Vote Intention t = ƒ (Party* t-1, Ideology* t-1, Cleavages, Issues) Eq. 5

16 14 The variables are measured as above, noting that the asterisk (*) indicates instrumental variables, and that the time subscripts, t-1 and t, indicate, respectively, whether the variable was measured in Wave 1 or Wave 2. See that Eq. 5 is already estimated in column 1, Table 1, while Eq. 3 is estimated in column 2, and Eq. 4 in column 3. One observes that ideology and party significantly and strongly affect each other, as expected. Further, one sees that ideology affects party at.93, a stronger affect than party on ideology, at.81. Thus, to the extent that there are indirect effects from party and ideology on the vote, they underline the dominance of ideology. The causal system among these variables, along with these path estimates, is sketched in Figure 1a. (Figure 1a about here) One criticism of these 2SLS estimates is that the instrumental variables are poor, because they are not constructed from variables that are truly exogenous (see the discussion in Woolridge, 2006, pp ). Considerable care was taken in the selection of the exogenous variables used to build these instruments, as was discussed above in footnote 4. However, in research practice, some choices are almost inevitably open to challenge. Therefore, we carried out further exogeneity tests. First, we built the instruments imposing a very strict exogeneity assumption, which meant only socio-demographic Wave 1 measures were employed, namely age, gender, education, income, occupation, work sector, and organizational membership. Then, we re-estimated Eq. 5. The absolute values of the slope coefficients for these ideology and party instruments are, respectively,.59 and.17, showing the continued dominance of ideology. Two more experimental analyses were run: one creating the instruments by adding

17 15 to these seven core socio-demographic variables the three retrospective evaluations of democracy, political interest, and the EU; another creating the instruments by adding to these seven core socio-demographic variables the 1995 vote. In both these analyses, ideology maintained its superior display of strength. Respective absolute values are 1.05 and.64,.61 and.43, thus sustaining the conclusions from our original instrumental variables. We choose to stay with these, as they have the additional virtue of being good instruments in another way; i.e., they predict the original endogenous explanatory variables better than the experimental alternatives. Another criticism of these Table 1 results is that the estimates rest on the assumption of interval data. The dependent variable of vote intention, while it has a seven-point scale, is more nearly ordinal than numeric. Therefore, we re-ran the analysis for the three equations (Eqs. 3-5) using ordered logit, which conforms better to the actual measurement level (Borooah, 2002). The results appear in Table 2, with the path diagram presented in Figure 1b. The essential findings of Table 1 continue to be supported: ideology has more influence over vote intention than party. If any thing, the dominance of ideology over party appears greater than it did in Figure 1a, since for direct effects,.90 >.55. (Table 2 and Figure 1b about here) It is possible to object to the ordinal assumption associated with the dependent variable above, on grounds that the vote choice is categorical and multidimensional. Therefore, we specified various binomial and multinomial logits. We first ran binomial logit estimations with the dependent variable as vote intention for each of the two main French parties, namely

18 16 the Socialist Party and the RPR. The results appear in the first two columns of Table 3. They confirm again the greater influence of left-right ideology over vote intention, everything else being equal. Look first at the likelihood of supporting the incumbent Socialist government versus any other party (column 1). Ideology has a logit coefficient of compared to only -.71 for party identification. It is also useful to look at the likelihood of legislative support for the RPR, given the designation of actual head of government was in transition during this electoral period (column 2). That is to say, the vote intention dependent variable is expressed after the presidential election of Chirac, but before the legislative election which would establish the parliamentary majority. The RPR vote estimation indicates that ideology dominates, with a logit coefficient for ideology of 1.44 compared to.46 for party identification. (Table 3 about here) Multinomial logit is another discrete choice estimation technique that can be used to examine the issue at hand. We incorporate the above categories, RPR, Socialist, Other parties into a three-way choice model. The baseline category is Socialist, since it was incumbent; thus, the basic voter choice posed in the model is staying with the Socialist government versus voting for the mainstream opposition, or for the extreme opposition. Precedent for this approach comes from Gschwend and Leuffen (2005), who pose the same type of three-way MNL dependent variable for the French case. The third and fourth columns of Table 3 report our multinomial logit estimates of first-round legislative vote intention in Again, this shows that ideology dominates party; respectively, 1.79 >.80,.75 >.48. Further, these results

19 17 are supported in a full seven-way multinomial logit estimation, which shows that ideology dominates party for all pairs of categories when either PS or RPR is the baseline category, with only one exception (choice of UDF over PS). (Results available from the authors upon request.) The foregoing analysis demonstrates the continued relative strength of ideology, as the left-hand side measurement assumptions are successively relaxed. However, an objection could be raised about the right-hand side assumptions as well. What happens when the interval assumption of the independent variables of ideology and party are similarly relaxed? This may be particularly important to do, in order not to stack the deck against the party variable. Therefore, we dummied up the variables P* and I*, producing six categorical variables each to capture party identification and ideological identification. The Extreme Left stands as the excluded category for each. (For details on the coding, see the footnote). [6] The basic idea is to re-run the core model of Table 3, substituting in these categorical party and ideology measures on the right-hand side. In Table 4, binomial logit equations of Socialist vote intention (column 1) and RPR vote intention (column 2) are reported. Note that the same controls are added to the party and ideology variables, which are now represented as dummies. In terms of effects, ideology is again superior, according to a comparison of the average dummy ideology coefficient to the average dummy party coefficient; respectively, those averages are 1.61 > 1.28 (Socialist baseline), 2.77 >.46 (RPR baseline). The same comparison for medians continues to give ideology more strength; respectively, 1.68 > 1.30, 3.29 >.38. Furthermore, one observes that, across the two models, ideology dummies have eight statistically significant coefficients,

20 18 while only three of the party dummies are statistically significant. These results make clear that ideology dominates party. (Table 4 about here) At this point, the more complex, three-way and seven-way multinomial logit models merit consideration. These models were run (results available upon request), but serious problems were encountered, because of the large number of dummies on the right- and lefthand sides of the equations. Specifically, there were grave difficulties from severe collinearity, extremely underpopulated cells and, in the case of the seven-way analysis, the need to estimate 102 coefficients simultaneously. Nevertheless, a general finding does emerge from these analyses. When separate models are run using only the ideology dummies, or the party dummies, respectively, the pseudo-r-squared for the ideology models always exceed the pseudo-r-squared of the party models. This suggests that our central hypothesis the ideology effect exceeds the party effect holds even under these conditions. Before drawing final conclusions, a simple, but telling, test is in order. How many respondents actually changed their party identification over the course of the study? Whereas the Wave 1 question about left-right ideological identification was not asked in later waves of the panel, the question about party identification was. We can thus assess the degree to which party identification changed between Waves 1 and 2, as well as the direction of those changes. In addition, we can determine whether those panel respondents who switched party identification between waves remained (or not) within the same ideological camp. [7]

21 19 Figure 2 summarizes those results. We can first observe that 37% of the panelists changed party identification between the two waves. About one third of these respondents are new (or mobilized) partisans, that is individuals who mentioned no party in Wave 1 but then declared being close to one in Wave 2. The remaining two thirds are respondents who actually switched party from one wave to the next. Within that group, we distinguish between those partisans who switched ideological sides or not. That is, we divide the French partisan space into two distinct ideological blocs, the left and the right [8], and see how many partisans actually went to the other side. It turns out that 141 panelists switched to a party belonging to the opposite bloc (perhaps from measurement error). [9] By way of contrast, 279 switched to a party ideologically close to the first one. In other words, fully two thirds of partisan switchers nevertheless stayed on the same ideological side. (Figure 2 about here) What prompts French voters to change party identification? The main reason appears to be the relative weakness of that social-psychological attachment (see also Chiche, Haegel and Tiberj, 2004). Table 5 first indicates that only 31% of panelists felt close or very close to a party (as measured in Wave 1). The table also shows that the weaker the attachment, the less stable the party identification. The proportion of respondents highly attached to a party is 40% among those who did not change party identification across waves, while this proportion drops to 28% among intra-bloc switchers and to 16% among extra-bloc switchers. In sum, the panel results indicate that, in France, party attachment is not as intense and stable as one might expect from a long-term social-psychological anchor. Left-right

22 20 ideology clearly appears as a stronger force structuring French political behavior. Crossing the ideological divide appears more exceptional than changing party identification, a conclusion that confirms a similar observation recently made about electoral volatility in France (see Swyngedouw, Boy and Mayer, 2000). (Table 5 about here) Discussion A number of reasons may explain why party identification is not as strong an anchor as ideological identification in France. One is the important fragmentation of the Fifth Republic s party system. Over the past half century, some French parties have been gradually marginalized (e.g., Communist Party, UDF) while several others have appeared on the political scene (e.g., Greens, Front National, extreme left parties). This makes for a party system in constant flux, both on the left and on the right. Another reason is the relative lack of meaning that major party labels generally have in France, as revealed through frequent changes in party names over the years. For instance, the Gaullist Rassemblement pour la République (RPR) started out as the Rassemblement du Peuple Français (RPF) in 1947, became the Union pour la Nouvelle République (UNR) in 1958 and then the Union des Démocrates pour la Cinquième République (UDR) in 1967 before adopting the label RPR in 1976; it has since been transformed into the Union pour un Mouvement Populaire (UMP) in November On the left, the Section Française de l Internationale Ouvrière (SFIO) became the Socialist Party in 1969, and was merged in 1971 with the Convention des Institutions Républicaines (CIR). Further, on the left, as well as the

23 21 right, many other smaller parties have come, gone, or transformed themselves (see Knapp, 2004, pp.6-7, for a fascinating and comprehensive summary table of the changing political party labels in France ). These characteristics of the French party system make it difficult for the electorate to develop a genuine long-term attachment to any particular party. In contrast, left-right ideology has always been a reliable guide to French voters for making sense of politics in their country, hence its superiority over party as a voting anchor. That being said, recent studies have hinted at the rise of new salient issue cleavages, such as immigration and insecurity, that are competing with the traditional left-right cleavage in structuring the political space in France (see for instance Chiche et al., 2000; Mayer and Tiberj, 2004; Evans and Mayer, 2005). While our own analysis indicates that those new issue dimensions are important to the vote, our study did not examine the extent to which these cleavages overlap with left and right, or whether they affect the traditional meanings of left and right in France. We believe it is to such questions that research on voting behavior in France should now turn. Summary and Conclusion The debate over the relative importance of ideology versus party for vote choice in France has been fought long and hard. Until now, it has seemed at a stalemate, with the most contemporary and methodologically advanced papers in sharp disagreement. But, this disagreement becomes reconcilable under new conditions, namely a research design which allows stronger causal inference.

24 22 Given the dynamics of the 2002 French Electoral Panel, the critical variables of party identification and ideological identification can be measured before the vote event, thus allowing the temporality of the causal chain to be taken into account. Further, using careful statistical techniques to disentangle the inherent reciprocal links, as well as the inevitable correlated errors, the party and ideology variables are exogenized. These variables serve as instruments to estimate the true effects, in a two-stage least squares or ordered logit estimation of a fully specified multi-equation model applied across the panel. The conclusion, borne out by further analyses employing binomial and multinomial logit techniques, is that ideological identification systematically outweighs party identification in shaping the vote. Of course, that conclusion may be restricted to the first-ballot legislative contest, and for the 2002 election in particular. But for that contest, the conclusion appears inescapable. Finally, with respect to the larger question of the relative importance of ideology and party in democratic polities beyond France, we believe our findings suggest two avenues for future research: 1) what is the relative electoral impact of ideological identification and party identification in the mass publics of other democracies? 2) what are the party system consequences of these differences in the relative weight of these two long-term anchors?

25 23 FOOTNOTES [1] In Wave 3, actual vote was asked, as opposed to vote intention, in Wave 2. Unfortunately, Wave 3 variables could not be effectively used in our panel analysis, because of the number of relevant missing items across the waves. Response rates also differed across the waves. For Wave 1, the final sample was actually based on two samples of approximately 2,000 each week, representing the population of registered voters. Out of these, 55% said they were willing to be reinterviewed for Wave 2. Of these, 1,822 actually submitted to the second interview, giving an effective response rate for Wave 2 panelists of 81%. (Extensive details on the sampling frame and the response rates are available in Cautrès and Jadot, 2004). [2] With open-ended party identification questions in France, very large numbers of distinct responses tend to be recorded (Lewis-Beck and Chlarson, 2002, p.512). Here the respondents are asked to select from a closed list of 13 different partisan categories. As can be seen, we reduce these 13 choices to 7 categories. These combinations are straightforward, either combining splinter parties to each other, or to its major ally. One coding decision that might raise a question is that of combining Hunters with Greens and other Ecologists. This was done on the grounds that they are all concerned primarily with environmental issues, as they conceive them. Knapp (2004, pp. 282, 285), in his definitive recent treatment of the French party system, makes this connection. He also notes that the electorate of the Chasse, Pêche, Nature et Traditions (CPNT) party, to give it its full name, is one-third left, one-third center, and one-third right. Still, its greater concerns over traditional issues do make any overlap with more modern environmentalist parties far from complete. Therefore, as an experiment, we excluded from our analysis those eleven panelists who identified themselves with the CPNT; this exclusion had virtually no effect on our results.

26 24 [3] A classic treatment of using exogenous variables in a panel survey, in order to create instrumental variables, appears in Fiorina s (1981, Appendix B) analysis of American National Election Studies data. He identifies as exogenous variables that are demographic variables and are retrospective evaluations. Two retrospective variables we use are the 1995 Chirac and Jospin vote variables, which are dummies indicating whether the respondent said they voted for one or the other candidate in the second round of the 1995 presidential election. Some might consider these variables as not being entirely exogenous. However, use of recollected political support as an instrumental variable is not an uncommon practice. For example, Fiorina (1981, Appendix B) used as exogenous recalled parents party identification and presidential popularity. We would argue that recalling 1995 vote when asked in 2002 has less endogeneity than either recalled parents party identification or presidential popularity. [4] To create the instruments, at first all twelve exogenous variables were regressed on the independent endogenous variables. However, as sometimes happens, a debilitating collinearity problem between P* t-1 and I* t-1 resulted (Berry, 1984, p.71). Once the Chirac variable was excluded from the construction of I* t-1, and the Jospin variable was excluded from the construction of P* t-1, the problem was solved. Further, it was found that not all the exogenous variables were necessary to maximize fit; therefore, for parsimony, only exogenous variables that managed a significant increment to the R-squared were used. For party identification (P t-1 ), that meant that the following six exogenous variables went into the creation of P* t-1 : democratic satisfaction, organizational membership, support for the EU, income, 1995 Chirac vote, and work sector (yielding an equation R-squared =.36). For ideological identification (I t-1 ), the useful exogenous variables were the following five: age, income, work sector, interest in politics, 1995 Jospin vote (yielding an equation R-squared =

27 25.37). The two instruments, I* t-1 and P* t-1, correlate at.67, not high enough to induce any collinearity problems. [5] The sample size (N = 878) for the vote equation in column 1, Table 1 shows considerable attrition, due to the presence of the social class variable, which about one-third of respondents did not answer. If, as an experiment, this social class variable is dropped and the analysis reran, the N jumps to 1,328 and our conclusions about party and ideology are unaffected. Indeed, in that specification, the dominance of ideology over party appears even stronger. [6] These dummies were constructed as follows. Cut-points were established for the seven categories of party (and ideology) by comparing the distributions of variables P and P* (and I and I*). For example, with P, 3.4% of the respondents identified with the Extrême Gauche. Therefore, the first 3.4% of the respondents in P* were coded as Extrême Gauche (the interval going from 2.99 to 3.14). The cut-points, then, are not at equal intervals of each other, but rather determined so as to maximize the resemblance between the two distributions. (The alternative, of equal intervals, fails because it allocates highly unrealistic shares to the parties; e.g., such a method would put 20% of the respondents as Extrême Gauche, and almost 40% as Communist). [7] See Chiche, Haegel and Tiberj (2004) for similar analyses of change in party identification between the different waves of the 2002 French Electoral Panel. Their conclusions are virtually identical to ours on that account. [8] See Andersen and Evans (2003). Grunberg and Schweisguth (1997) argue for a tripartion of the French political space; note that our results remain roughly similar when we distinguish three blocs (as they suggest): left, moderate right and extreme right.

28 26 [9] These 141 extra-bloc switchers equal 8% of the relevant sample, equivalent to a normal estimate of measurement error at 8%. While it would be highly unusual for all the measurement error to be concentrated within this group, there is undoubtedly some.

29 27 References Andersen, Robert, and Jocelyn A.J. Evans Values, Cleavages and Party Choice in France, , French Politics 1: Berry, William D Nonrecursive Causal Models, Beverly Hills: Sage. Borooah, Vani Kant Logit and Probit: Ordered and Multinomial Models, Thousand Oaks: Sage. Boy, Daniel, and Nonna Mayer The Changing French Voter, in Daniel Boy and Nonna Mayer (eds.), The French Voter Decides, Ann Arbor: University of Michigan Press, p Budge, Ian, Ivor Crewe, and Dennis Farlie (eds.) Party Identification and Beyond: Representations of Voting and Party Competition, London: Wiley. Cautrès, Bruno, and Anne Jadot Panel, mode d emploi, in Bruno Cautrès and Nonna Mayer (eds.), Le nouveau désordre électoral: les leçons du 21 avril 2002, Paris: Presses de Sciences Po, p Cautrès, Bruno, Nonna Mayer, and Annie-Claude Salomon The French Electoral Panel 2002, French Politics 1: Chiche, Jean, Brigitte Le Roux, Pascal Perrineau, and Henry Rouanet L espace politique des électeurs français à la fin des années 1990, Revue française de science politique 50: Chiche, Jean, Florence Haegel, and Vincent Tiberj Érosion et mobilité partisanes, in Bruno Cautrès and Nonna Mayer (eds.), Le nouveau désordre électoral: les leçons du 21 avril 2002, Paris: Presses de Sciences Po, p Converse, Philip E., and Roy Pierce Political Representation in France, Cambridge: Belknap-Harvard University Press. Dalton, Russell J., Scott Flanagan, and Paul Allen Beck (eds.) Electoral Change in Advanced Industrial Democracies: Realignment or Dealignment?, Princeton: Princeton University Press. Elgie, Robert The Institutional Logics of Presidential Elections, in Robert Elgie (ed.), Electing the French President: The 1995 Presidential Election, New York: St. Martin Press, p Evans, Jocelyn A.J Ideology and Party Identification: A Normalisation of French Voting Anchors?, in Michael S. Lewis-Beck (ed.), The French Voter: Before and After the 2002 Elections, Basingstoke: Palgrave Macmillan, p

30 28 Evans, Jocelyn A.J., and Nonna Mayer Electorates, New Cleavages and Social Structures, in Alistair Cole, Patrick Le Galès and Jonah Levy (eds.), Developments in French Politics 3, Basingstoke: Palgrave Macmillan, p Finkel, Steven E Causal Analysis with Panel Data, Thousand Oaks: Sage. Fiorina, Morris P Retrospective Voting in American National Elections, New Haven: Yale University Press. Fleury, Christopher J., and Michael S. Lewis-Beck Anchoring the French Voter: Ideology versus Party, Journal of Politics 55: Franklin, Mark N., Thomas T. Mackie, and Henry Valen (eds.) Electoral Change: Responses to Evolving Social and Attitudinal Structures in Western Countries, Cambridge: Cambridge University Press. Goguel, François Les élections législatives des 23 et 30 juin 1968, Revue française de science politique 18: Grunberg, Gérard, et Étienne Schweisguth Vers une tripartition de l espace politique, in Daniel Boy and Nonna Mayer (eds.), L électeur a ses raisons, Paris: Presses de Sciences Po, p Gschwend, Thomas, and Dirk Leuffen Divided We Stand Unified We Govern? Cohabitation and Regime Voting in the 2002 French Elections, British Journal of Political Science 35: Haegel, Florence Le lien partisan, in Daniel Boy and Nonna Mayer (eds.), L électeur français en questions, Paris: Presses de Sciences Po, p Knapp, Andrew Parties and the Party System in France: A Disconnected Democracy?, Basingstoke: Palgrave Macmillan. Lewis-Beck, Michael S Cross-National Election Surveys: A French Pre-Test, Electoral Studies 15: Lewis-Beck, Michael S French Election Theories and the 2002 Results: An Introduction, in Michael S. Lewis-Beck (ed.), The French Voter: Before and After the 2002 Elections, Basingstoke: Palgrave Macmillan, p Lewis-Beck, Michael S., and Kevin Chlarson Party, Ideology, Institutions and the 1995 French Presidential Election, British Journal of Political Science 32: Lewis-Beck, Michael S., and Glenn E. Mitchell, II French Electoral Theory: The National Front Test, Electoral Studies 12: Lewis-Beck, Michael S., and Andrew Skalaban France, in Mark N. Franklin, Thomas T. Mackie and Henry Valen (eds.), Electoral Change: Responses to Evolving Social and