This is the author's version of the work. Please cite as:

    Arzheimer, Kai and Jocelyn Evans. "Bread and butter à la française: Multiparty forecasts of the French legislative vote 1981-2007." International Journal of Forecasting 26 (2010): 19–31. doi:10.1016/j.ijforecast.2009.05.025
    [BibTeX] [HTML] [DATA]
    @Article{arzheimer-evans-2010,
    author = {Arzheimer, Kai and Evans, Jocelyn},
    title = {Bread and butter à la française: Multiparty forecasts of the French legislative vote 1981-2007},
    journal = {International Journal of Forecasting},
    year = 2010,
    volume = 26,
    pages = {19--31},
    keywords = {voting, france, subnational},
    html = {https://www.kai-arzheimer.com/paper/bread-butter-a-la-francaise-forecasts-french-legislative-vote-regional-economic-conditions/},
    data = {https://doi.org/10.7910/DVN/CYB7QZ},
    doi = {10.1016/j.ijforecast.2009.05.025}
    }

Bread and butter à la française: forecasts of the French legislative vote from regional economic conditions

Introduction

Electoral forecasting is a longstanding sub-discipline of psephological research, but one which has concentrated principally on Anglo-Saxon systems, and in particular two-party systems with a straight incumbent / opposition trade-off in votes. Multiparty systems with more complex interrelationships between party vote shares have remained a minority field in forecasting work, and certainly theoretical and methodological advances in forecasting overall vote outcomes are much less common.

The simple association between economic conditions and level of incumbent support, either through election or public opinion polling, is the cornerstone of the vote-popularity (VP) function literature in its economic guise (Lewis-Beck 1988), and its applicability has been shown to be clearer in single-party government (two-party) systems than in multiparty coalition settings (Powell and Whitten 1993). Multiparty systems present a more complicated – confused even – situation of accountability for economic downturns, through coalition responsibility obscuring a target for dissatisfied voters. Moreover, in the French case, the semi-presidential “bicephalous” executive has prompted the question ‘Who’s the chef’ (Lewis-Beck 1997) with regard to allocation of economic responsibility by voters between Prime Minister and President.

However, a more conceptual problem with regard to multiparty systems is how satisfactorily to construct an economic forecasting model that allows prediction of a set of inter-related votes across multiple parties. From the Anglo-Saxon roots of vote forecasting, the zero-sum incumbent / opposition approach which lends itself well to simple linear regression modelling is inapplicable to a model which wishes to go beyond a simple bloc prediction or restrict itself to a single party prediction. This paper sets out to provide a constrained ‘seemingly unrelated regression’ (SUR) model predicting all votes in French legislative elections between 1981 and 2002, and uses this to forecast the 2007 legislative election result.

The French case

Amongst non-Anglo Saxon cases, France has increasingly become a country of interest to electoral forecasters. There is a growing tradition of election forecasting of incumbent vote (e.g. Jérôme et al 1999; Jérôme et al 2003; Auberger and Dubois 2005), as well as estimates of third-party performance (Jérôme and Jérôme-Speziari 2003; Auberger 2008; Evans and Ivaldi 2008) and second-order elections (Jérôme and Jérôme-Speziari 2004; Auberger 2005). The majority of this work provides discrete estimates for individual parties or blocs. There is little attempt to find a single model predicting distribution of the totality of vote across all parties, however. Given the archetypal multipartism present in the French system, approaches based upon the two-party adversarial models used in the United States and the United Kingdom (over-) simplify the relative party scores because of their underlying zero-sum logic.

Moreover, in common with all elections, but of specific interest in France, it is important to take into account not just voting but also non-voting. In France, abstention is regarded as ‘wasted’, as it is not factored into the valid vote upon which outcomes such as progression to second rounds are based. However, its effect in denying political parties the support of their core electorate has been perceived as a significant influence on election outcomes, not least in the 2002 presidential election where the Socialist candidate failed to reach the run-off stage due to relatively poor turnout amongst grassroots Socialist supporters (Shields 2005). Abstention takes on even greater importance in legislative elections which follow presidential races, as losing candidates in the presidential election generally see their parties lose even more support in the subsequent parliamentary race.1

In many ways, electoral patterns in France since the mid-1970s have encouraged the use of government / opposition models, given the establishment of the so-called ‘bipolar quadrille’ of Left and Right blocs, and the alternation between the two at ever legislative election from 1981 to 2007 – ‘hyper-alternance’, as it has been termed (Evans and Ivaldi 2002). However, the importance of third parties in disrupting this stable two-bloc pattern, from the late 1980s ‘til 2002, means that a unified model providing a set of predictions across all party groups is increasingly important. Even in the years of the moderate two-bloc bipolar quadrille, vote distribution amongst the two Right-wing parties (RPR and UDF) and the Socialists and Communists could not be assumed to be a simple sum of the two bloc-members on either side. Furthermore, the rise of the Extreme Right FN, and particularly of its presidential candidate Jean-Marie Le Pen, upset any premise whatsoever based upon a straight Left / Right contrast.2 Indeed, looking again at the 2002 presidential elections, the failure of the Socialist candidate to reach the second round due to high abstention was compounded by a strong showing by Le Pen, with the consequent unexpected progression of the Extreme Right candidate precisely into the ballottage.

As a case for third-party prediction, then, France and the FN / Le Pen have become compelling. To what extent has this party’s performance been predictable, and was the 2002 result an aberration or consistent with a possible model of Extreme Right voting? Given the small number of cases with which to construct a model, the notions of outliers is somewhat problematic in analytical terms, but focus on attempting to predict the FN and Le Pen’s performance in 2007 suggested that in fact the success would be repeated. The majority of predictions for 2007 saw no reason to expect anything other than a strong performance by the Extreme Right (Lewis-Beck et al, 2008). The majority of predictions were consequently wrong, with Le Pen and then the FN suffering a relative electoral rout. Models predicated upon conditions of success in 2002 fell apart in 2007.

However, despite the real-world necessity of focusing on this single candidate / party’s vote, methodologically there are issues with simply attempting to predict a single party’s vote in a multiparty setting. Whilst it may in practice be possible to predict a single party’s vote3, this does not take into account provenance of the vote or consider how the other parties’ votes relate to the outcome of the party in question. Parties do not contest elections in isolation, with votes varying purely in relation to their own electoral performance: performance of other parties must have an effect. If one wishes to look at the entirety of the vote distribution, and how this can be predicted, multiple equations may provide counter-intuitive results such as vote scores summing to more than 100%. A constrained model producing scores for all party alternatives is a more appropriate approach.

Before moving to the modelling process itself, we can briefly state the explanatory variables which we include in the model. Our predictors, which we operationalise in the following section, are from the foundations of economic forecast modelling. In France, evidence has been found of relatively sophisticated economic voting, with voters able to allocate responsibility for economic management amongst incumbents (Lewis-Beck, 1997). More broadly, the state of the national economy “matters” in politics (Auberger, 2004) and indeed influences second-order as well as legislative elections (Jérôme and Jérôme-Speziari, 2004). Our predictors present no surprises. First, we use unemployment as a key indicator of economic stability, and one closely associated with shifts in incumbent party support. Previous research in France has shown clearly shown decreases in incumbent party support when unemployment increases (Jérôme et al, 1999; Auberger and Dubois, 2005). Equally, a large literature exists linking Extreme Right party voting with this variable (Jackman and Volpert 1996, Lubbers et al. 2001, Golder 2003).4 Secondly, we look at the effect of GDP as a broader independent measure of economic success and stability. Again, this is a staple of the classic VP-function literature, though is less analysed in Extreme Right voting studies, not least because of a less clear link between a broad macro-economic indicator and extremist support than with unemployment, which taps into Extreme Right appeals to direct social precariousness. In French legislative election forecasts, the role of national economic growth is less common as a predictor, but where used, has been demonstrated as relating positively with incumbent support (Auberger and Dubois, 2005). Lastly, we also include a political variable, Left incumbency which we combine in an interaction with unemployment, to explore whether there is any difference in the way voters penalise, or indeed reward, for changes in economic fortunes under different parties – an economic ‘issue ownership’ effect which has been found in previous research using France as a case (Whitten and Palmer 1999).

Data and methods1

Even if we considerably simplify the number and structures of choices voters face in France, there are at least six different alternatives available to them: they may abstain (the reference category) or vote for the Communists, the Moderate Left (Socialist Party and allies, including the Parti Radical de Gauche), the Moderate Right (Gaullist RPR / UMP, UDF and Nouveau Centre), the Extreme Right (the Front National and splinter parties like the Mouvement National Républicain), or any other party (chiefly the Extreme Left groupuscules and the Greens, but also including regional parties and spoilt / blank ballots).5

Principally, then, we have six interrelated vote shares at the departmental level. This setup renders linear regression – still the workhorse of political science research – unfeasible. Linear regression requires a single unbounded dependent variable with errors that are distributed identically and independently. Vote shares, by contrast, are bounded to the interval [0;100]. Even if we ignore this boundary problem and try to model the six shares separately, vote shares are dependent by definition because they must always sum up to 100: if one party (say the Front National) does particularly well, the pool of voters available for other parties becomes smaller or turnout increases.

Two solutions to these problems have been suggested in the literature: Katz and King (1999) propose a complex likelihood estimator for multiparty data that is exact. Unfortunately, their method requires defining and programming one’s own likelihood functions and becomes computationally extremely demanding for more than three parties. A more viable three-step strategy was outlined by Tomz et al. (2002). First, an arbitrary option is chosen as the reference category, and log-ratios are calculated, which removes the boundary problem. For instance, in Paris (75), the abstention rate was 38.9 per cent in 1988. In the same election, 9.9 per cent of those eligible to vote voted for the FN, 32.3 per cent for the moderate left, and so on. Therefore, the log-ratio for the FN = ln(9.9/38.9) = -1.37, and the log-ratio for the Moderate Left is ln(32.3/38.9) = -0.19.

Second, the effects of the independent variables on the set of transformed variables are estimated via seemingly unrelated regression (SUR), a multi-equation version of linear regression. SUR allows for errors that are correlated across equations. Third, the predictions from this model are converted back to the original scale (percentages of those eligible to vote). While the Katz-King method is more accurate in theory, the extensive analyses by Tomz et al. show that their simpler approach performs equally well in practice.

There is, however, one additional complication. While Tomz et al. model single elections, the structure of our data is time-series cross-sectional, i.e. we have (rather short) time-series of results for a “panel” of 96 departments, each with its own socio-structural composition and tradition that distinguishes it from the other districts. To incorporate this structure into our model, we specify a fixed effect for each department that reflects its voting history, political tradition and other more or less constant (over the course of two and a half decades) factors which affect the outcome of a given election. This implies that our model focuses on variation over time within departments while time-invariant differences between departments (that cannot possibly be the result of our dynamic independent variables) are taken for granted.

Another implication of the time-series aspect of our data is that errors within a department will most probably track over time. A large part of this variation will be soaked up by the fixed effects, but as an additional safeguard, we specify robust (Huber-White) estimators for the variance that account for this dependency, thereby yielding conservative standard errors.6

Our dependent variables are the vote shares and turnout rates at the departmental level from the first rounds of the legislative elections from 1981 to 2002 in Metropolitan France.7 For 1981 to 1997, these figures were compiled by Caramani (2000). Results for 2002 and 2007 (for checking the accuracy of our predictions) were taken from the Centre de Données Socio-Politiques’s website (http://cdsp.sciences-po.fr/AfficheElec.php). Unemployment figures on the departmental level come from quarterly figures collected by INSEE (http://www.bdm.insee.fr/bdm2/do/accueil/AccueilAppli). Information on GDP growth (which is collected at the national level and therefore constant across departments in any given election year) comes from Eurostat structural indicators (http://epp.eurostat.ec.europa.eu). Data are not lagged, i.e. we use unemployment rates and GDP growth for the year in which the respective election was held. We tested for possible cross-correlation between unemployment and GDP growth, and found no evidence of this.8

Analysis

[Table 1 about here]

We first examine the overall model fit statistics with a nesting to examine the degree to which political tradition at departmental level accounts for party bloc support, and the improvement in fit that the economic and political indicators provide (Table 1). Overall, it is clear that the additional forecasting variables provide a substantial improvement in fit, as evidenced by the shrinking of the Bayesian Information Criterion (BIC). However, the differential change in R2 from the fixed effects baseline indicates varying political traditions as well as responsiveness to economic variation. Broadly speaking, Left parties, and particularly the Communists with almost half the variation in their vote accounted for by departmental effect, enjoy stronger baseline support in traditional Left-oriented areas (and conversely consistently low support in non-leftist departments). The Moderate Right sees around one-third of variation accounted for by this. Conversely, the Extreme Right and other parties are not determined by localised consistency, with much smaller R2.

The move to the full model shows a much different pattern across the party blocs. The Moderate blocs and the other parties see a relative improvement in fit once economic indicators are taken into account,9 but not to the same extent as the Extreme Right, which sees economic indicators improve the model fit very substantially. In accounting for Extreme Right support, then, macro-economic indicators matter more than any longer term departmental tradition. At the other end of the spectrum, the Extreme Left Communists see a much more muted improvement, particularly in comparison to the fixed effects baseline. Political tradition in its ‘bastions’ has historically mattered more for the PCF than it does for other parties – something which has been noted elsewhere in the literature on Communist support in France and elsewhere (Bell and Criddle, 1994).

[Table 2 about here]

We now turn to the parameter estimates of the regression model to interpret the above changes in more detail (Table 2). There are both intuitive and less clear effects. Recall, firstly, that all reported estimates in SUR are contrasts, in this case with the abstention reference category. Consequently, parameters should not be read as absolute values for each party grouping, but rather relative effects. Rises in unemployment see a drop in support for all the mainstream parties, relative to abstention. Conversely, with only support for the Extreme Right and oOther parties increasesing. Equally, given the contrast category in the logit, unemployment appears to reduce participation. Abstention and votes for marginal and extreme parties benefits, then, from downturns in the jobs market. Similarly, Aa rise in national GDP will increase participation, benefiting both the mainstream parties and as well as the Extreme Right party, but ERP – less easily explicable – and disfavouring marginal parties. Unlike any other party, then, the FN in France appears to benefit from both economic upturns and downturns, but dependent upon the area of the economy.

The Left incumbency effect, as might be expected, penalises Left-wing parties, but (in particular the Communists. When the Left has been in power, supporters of these parties increasingly abstain, in net terms.) Bbut it also appears to penalise the Moderate Right in a similar way. We can only speculate that this may be a general participatory malaise which depresses turnout for all mainstream parties in the ‘aftermath’ of a Left-wing government – in the cases we include in this model, Left-wing governments generally performed singularly badly in economic terms., Apparently, however, although the mainstream Right opposition appears not to have inspired particular confidence as an alternative!and thus support as a replacement.10

Lastly, we turn to the interaction effect between unemployment and Left-wing incumbency effectively indicates the ‘premium’ that the latter imposes on the former’s effect. The interaction does not exhibit what we might expect it to. If we hypothesise ‘issue ownership’ of unemployment as a policy-area associated with the Left, we might expect support for Moderate Left parties to decline more heavily when they have been in government and unemployment has increased. Under Socialist governments, the electoral penalty for rises in unemployment on mainstream parties is less pronouncedower than under Right-wing governments (i.e. when the interaction effect is effectively zero, given the Right-wing reference in the incumbency variable). The same is true for Right-wing support as well, as indicated by the identical parameter estimate., and Cconversely, the Extreme Right and other parties do less well, losing more support to abstention when unemployment rises under a Left-wing government. In that respect, there does not seem to be evidence of ‘issue ownership’ of unemployment by Left governments which sees them punished more strongly than Right governments by rising unemployment.

[Figure 1 about here]
  • Maybe cut the interpretation of the parameters a bit, point out that the parameter-based interpretation of a multi-nomial model can be confusing
  • We focus on the median growth level of 1.9 per cent to keep things simple and because growth does not seem to make too much of a difference for most choices at most levels of unemployment and incumbency.
  • One of the scripts in the replication archive is easily adjusted to estimate vote shares for different growth levels.
  • That being said, a decline in the economy seems to boost the prospects of the “other” choices (though empirical support for this effect is thin) and reducing support for the FN, while strong growth seems to improve the prospects of the moderate left (if the economy is doing well, we can even afford a socialist government!) and, to a lesser degree, the prospects of the Communists.
  • Both the major left and the major right pay a price for governing (there must be a reference for this or a similar phrase which I cannot remember): ceteris paribus, support for them is lower when they are in government cf. top-left and top-centre graphs
  • This effect reflects on the electoral fortunes of the extreme parties within the blocks: the FN does better under a leftist government, while the Communists are better off under a right-wing government (though this effect is diminished when unemployment is very high). Cf. bottom-left and bottom-centre.
  • The moderate left is punished severely for high unemployment, even if they are not in government, whereas the moderate right can get away with moderate increases in unemployment. Even if unemployment is very high, the moderate right performs better than the moderate left (cf. top left vs. top centre)
  • Unemployment boosts FN voting, especially under a leftist government. (bottom left)
  • The prospects for “other”parties increase with unemployment; dramatically so if there is a moderate right government. (bottom right)
  • Non-voting increases (i.e. turnout goes down) with unemployment under a right-wing government. Non-voting goes down (i.e. turnout goes up) with higher unemployment rates if there is an incumbent left government (top-right). The FN, communists and other parties would benefit from this increase in turnout.
[Table 3 about here]

We move now to the predictive efficacy of the model for the 2007 legislative elections (Table 3). Recall that the vote share indicated includes abstention and spoilt/blank ballots. We therefore provide a table which includes not only the notional vote shares for 2007 as a proportion of registered voters under our predictive model, but also predicted and observed election outcomes as a proportion of votes cast.11 At first glance, the prediction is disappointing. All mainstream party scores are significantly over-estimated, with a concomitant under-estimation of the apparently ‘protest’ or ‘dissatisfaction’ votes. In terms of precision, then, our model does not pass with flying colours. That said, there are some positive aspects. Firstly, the model does correctly predict the winner of the election. Whilst this is less important for a relative majority system such as France than a two-horse race, it still passes the test of ‘spotting the winner’ – not always the case with ex ante forecasts (Lewis-Beck et al 2008).

Similarly the model actually under-estimates the Extreme Right score. As we noted in the overview of the French context, a model able to predict with some degree of accuracy the Extreme Right vote in France has since 2002 been an increasing focus of the forecasting literature. However, precisely given the exaggerated success – in qualitative terms – given to Le Pen’s performance in the 2002 presidentials, ex ante predictions generally exaggerated – in quantitative terms – the likely Le Pen / FN score in 2007. Yet, a simple economic model plus incumbency control in fact gives a very respectable estimate.

Comparing votes as a proportion of registered voters and as a proportion of votes cast,12 a clear failing in the model is that it fails to predict that, even in terms of first-round votes, the Moderate Right grouping wins an absolute majority of votes. Clearly, this does not factor in seat distribution, which follows from the run-off results (although in the 2007 National Assembly election, the Moderate Right UMP group finished with 317 seats or 54.9% of the 577 total – very close to the first-round proportions).13 However, in a system dominated by the notion of majoritarianism, and the lowest 50%+1 mandate of its president, the psychological effect of surpassing 50% of votes cast at the first round is significant, and one which we would hope to pick up in a robust forecasting model.

The overall efficacy of the model in predicting 2007 voting levels in the election is clearly far from perfect. To try to identify issues with the model, we run some basic diagnostic tests on the model data (Appendix 1). Looking at outliers in each model, a small number of outliers stand out with high studentized residuals. In particular the Moderate Left contrast in Corse du Sud (Southern Corsica) has consistently high residuals in all years except 1988 and 1993, due to the localised party system including Corsican nationalist and local mainstream party variants. Due to the low leverage of the residuals, however, the removal of this department has little effect on the model. Similarly, for the ‘other’ parties, there are a number of underestimations in the 1980s elections for the Central South-West region and some North-East border areas. In many cases, this is due to so-called ‘majority’ candidates (i.e. candidates who explicitly support the President, and who were therefore supported by the official presidential party) who would in reality be more appropriately seen as Moderate Left, or in the case of the North-East, independent Right-wing candidates. For the Communists in 2002, a degree of under-estimation is present in a number of southern or Midi departments due to the disproportionately strong support from the working class, to an extent significantly beyond the political tradition of these areas.

Looking at studentized residuals for the Moderate Left and Communists party groups in 2007, 50 of the 96 departments were significantly over-estimated (cut-off for residual > 2.0) for the Moderate Left, and a massive 90 out of the 96 departments were over-estimated for the Communists. Overall, then, case diagnostics reveal little that indicates why the 2007 forecasts erred from the pattern set by 1981 – 2002, beyond the Left in toto performing much worse than previous patterns would suggest.

Discussion

Given the demands of using a constrained model in a complex multiparty setting with across-time variation in coalitions and party affiliation, the SUR approach provides a relatively robust model of French legislative election outcomes. Out-of-sample forecasting of the elections between 1981 and 2002 are on average quite strong, and concomitantly the parameter estimates for these models are generally consistent with what one would expect from the economic and VP-function literature. Inevitably, however, the panel nature of the data mean that changes in electoral coalition, mutual desistment pacts specific to individual elections and localised micro-party systems introduce bias into the model which it is difficult to correct. In an attempt to identify broad patterns economic effects on party support across time, such specificities will always explicitly influence predictions at the meso-level, and implicitly do so at the macro-level.

Of greater concern, perhaps, is the errant ex ante forecast of the 2007 result. Again, real life intervenes. In the French case, the key intervention which affects the legislative election outcome is the preceding presidential election. For four of our seven elections, a presidential race will have largely resolved the legislative outcome one month before the latter election. Yet, in 2007, there is good reason to believe that the resolution will have been more influential than ever. What had been expected to be a close presidential race, up until a relatively short time before the election, materialised into a comprehensive Right-wing victory over an unconvincing and divided Left, and the inevitable victory of Nicolas Sarkozy over the Socialist Ségolène Royal in the run-off. As a result, the legislative elections effectively became a confirmatory election to install President Sarkozy’s government, rather than a true electoral competition.

Similarly, the Extreme Right presidential candidate, Jean-Marie Le Pen, who had surpassed expectations in 2002, experienced a major downturn in electoral fortune in the 2007 presidentials. As with the mainstream rivalry, the legislative elections simply confirmed the electoral collapse of the Extreme Right with one of the worst national election scores ever for the FN. In a mirror image of the ‘unforeseeable’ success of Le Pen in 2002, so the crushing defeat was equally unforeseeable, at least in most of the predictive models prior to the election (Lewis-Beck et al 2008). Yet, as our model has shown, within a constrained model accounting for all party votes, the result was far from unforeseeable, and indeed this model could have been run a significant period before the legislative elections.14 Whilst the forecast is not precise, the prediction is not significantly outside the standard error of the model. To that extent, then, the 2007 performance was within the realms of prediction.

Whilst some additional refinement of party codings could potentially improve the model forecasts incrementally, the six-fold categorisation of the dependent variable does set the bar extremely high for a forecast model across 26 years of an extremely dynamic system. Undoubtedly, the SUR approach we have adopted is more robust than a simple linear regression approach for multiparty data. Yet, somewhat perversely, the evolution which prompted us to use this technique – namely, the rise in relevance in 2002 of a range of third parties and their presidential candidates – precisely seems to have been reversed in 2007. Looking back to Table 3 and the actual results, it is clear that the mainstream Moderate Left and Moderate Right blocs account for the vast proportion of the valid vote, with a noticeable shift back towards a two-bloc and indeed two-party logic (Grunberg and Haegel, 2008). In that sense, the more traditional incumbent / opposition approaches to electoral forecasting may well be increasingly appropriate once more for the French case. Whether a simplistic but effective forecasting outweighs a more nuanced but unstable approach is not a debate which can be resolved here. However, the latter approach which we have used here undoubtedly reveals a greater coherence and predictability to all types of party in the French system than perhaps previous research has been able.

References

Andersen, R. and J. Evans (2005) ‘The stability of French political space, 1988-2002’ in French Politics, 3:3, 282-301.

Anderson, C. (2000) ‘Economic voting and political context: a comparative perspective’, Electoral Studies, 19: 151-170.

Arzheimer, K. and Carter, E. (2006) ‘Political Opportunity Structures and Right-Wing Extremist Party Success’, European Journal of Political Research, 45, 419–443.

Auberger, A. (2004) ‘Les fonctions de vote: un survol de la littérature’, l’Actualité ÉconomiqueRevue d’Analyse Économique, 80: 95–107.

Auberger, A. (2005) “Forecasts of the 2004 French European election”, Swiss Political Science Review, 11, 61-78.

Auberger, A. (2008) “The National Front Vote and Turnout in the French Presidential Elections”, French Politics, 6:1, 94-100.

Auberger, A. and E. Dubois (2005) “The influence of local and national economic conditions on French Legislative elections”, Public Choice, 125, 363-383.

Bell, D. and B. Criddle (1994) The French Communist Party in the Fifth Republic, Oxford, Clarendon.

Caramani, D. (2000) Elections in Western Europe since 1815 : Electoral Results by Constituencies, Basingstoke, Macmillan.

Dubois, É. and C. Fauvelle-Aymar (2004) ‘Vote functions in France and the 2002 election forecast’, in M.S. Lewis-Beck (ed.) The French Voter: Before and After the 2002 Elections, Basingstoke, Palgrave Macmillan, 205–230.

Evans, J. and G. Ivaldi (2002) ‘Les dynamiques électorales de l’extrême-droite européenne’ in Revue Politique et Parlementaire, 1019, July-August, 67-83.

Evans, J. and G. Ivaldi (2008) ‘Forecasting the Extreme Right vote in France (1984-2007)’, French Politics, 6:2, 137-151.

Golder, M. (2003) ‘Explaining Variation in the Success of Extreme Right Parties in Western Europe’, Comparative Political Studies, 36, 432–466.

Grunberg, G. and F. Haegel (2008) La France vers le bipartisme?, Paris, les Nouveaux Débats (Presses de Sciences Po).

Grunberg, G. and E. Schweisguth (2003) ‘French political space: two, three or four blocs?’, French Politics, 1:3, 331-348.

Jackman, R. W. and K. Volpert (1996) ‘Conditions favouring parties of the Extreme Right in Western Europe’, British Journal of Political Science, 26, 501–521.

Jérôme B., V. Jérôme and M. Lewis-Beck (1999) “Polls fail in France: Forecasts of the 1997 legislative election”, International Journal of Forecasting 15: 163–174.

Jérôme, B. and V. Jérôme-Speziari (2003) “A Le Pen vote function for the 2002 Presidential election: a way to reduce uncertainty”, French Politics, 1:2, 247-251.

Jérôme, B. and V. Jérôme-Speziari (2004) “The 2004 French regional elections: politico-economic factors of a nationalized local ballot”, French Politics, 3:2, 142-163

Jérôme, B., V. Jérôme-Speziari and M. Lewis-Beck (2003) “Reordering the French election calendar: forecasting the consequences for 2002”, European Journal of Political Research, 42, 425-440.

Katz, J. and G. King (1999) “A statistical model for multiparty electoral data”, in American Political Science Review, 93:1, 15–32.

Knigge, P. (1998) “The ecological correlates of right-wing extremism in Western Europe”, European Journal of Political Research, 34:2, 249-79.

Lewis-Beck, M. (1988) Economic and Elections, Ann Arbor: University of Michigan Press.

Lewis-Beck, M. (1997) “Who’s the chef? Economic voting under a dual executive”, European Journal of Political Research, 31: 315-325.

Lewis-Beck, M., E. Bélanger and C. Fauvelle-Aymar (2008) “Forecasting the 2007 French Presidential Election: Ségolène Royal and the Iowa Model”, French Politics (2008) 6, 106–115.

Lubbers, M. and Scheepers, P. (2001) ‘Explaining the trend in Extreme Right-wing voting. Germany 1989-1998’, European Sociological Review, 17, 431–449.

Nannestad, Peter, and Martin Paldam (1994) ‘The VP-function – a survey of the literature on vote and popularity functions after 25 years’, Public Choice 79: 213-45.

Powell, G. Bingham and Guy Whitten (1993) ‘A cross-national analysis of economic voting: taking account of the political context’, American Journal of Political Science, 37: 391-414.

Shields, J. (2005) “Political Representation in France: a crisis of democracy”, Parliamentary Affairs 59:1,118-137.

Tomz, M., J. Tucker and J. Wittenberg (2002) “An easy and accurate regression model for multiparty electoral data”, Political Analysis, 10:1, 66–83.

Whitten, G. and H. Palmer (1999) ‘Cross-national analyses of economic voting’, Electoral Studies, 18: 49-67.

Table 1 Nested model comparisons, fixed effects and full model

forecasts (1981-2002)

Model

ll(null)

ll(model)

df

R2

BIC

Moderate Left: Fixed Effects

-394.22

-248.17

96

0.40

1106.54

Moderate Left: Full Model

-121.87

100

0.61

879.36

Moderate Right: Fixed Effects

-244.49

-135.97

96

0.31

882.12

Moderate Right: Full Model

-301.37

100

0.57

641.64

Communists: Fixed Effects

-633.09

-439.15

96

0.49

1487.98

Communists: Full Model

-374.60

100

0.59

1384.30

Extreme Right: Fixed Effects

-889.66

-850.12

96

0.14

2301.53

Extreme Right: Full Model

-643.54

100

0.61

1913.42

Other: Fixed Effects

-1445.50

-1414.72

96

0.11

3435.89

Other: Full Model

-1320.19

100

0.36

3272.09

Table 2 Parameter estimates for full model forecasts (1981-2002)

Comm

ML

MR

ER

Other

Unemployment (U)

-0.03 (.01)

-0.10 (.01)

-0.06 (.01)

0.53 (.03)

0.63 (.09)

Left incumbency (I)

-1.69 (.10)

-0.66 (.12)

-0.36 (.10)

3.14 (.35)

5.24 (1.04)

ΔGDP

0.00 (.01)

0.12 (.01)

0.05 (.01)

0.29 (.03)

-0.79 (.09)

U*I

0.14 (.01)

0.08 (.01)

0.08 (.01)

-0.15 (.03)

-0.46 (.10)

R2

.59

.61

.57

.61

.36

Root MSE

.51

.33

.27

.92

2.90

– All contrasts with abstention baseline

– Departmental dummies not shown

– All coefficients significant at .05 or lower, except Comm ΔGDP

Table 3 Predicted and observed vote outcomes in 2007 (% of total

registered voters and of votes cast)

% registered voters

% votes cast

Observed

Predicted

Observed

Predicted

Communists

2.7

8.8

4.4

13.3

Moderate Left

16.9

26.6

27.7

40.1

Moderate Right

32.6

29.0

53.4

43.7

Extreme Right

2.9

1.2

4.8

1.8

Other

5.8

0.9

9.5

1.4

Abstention

39.0

33.7

Bread and butter à la française: forecasts of the French legislative vote from regional economic conditions 1

Figure 1: Unemployment and Voting in France

Appendix 1 Outlier Studentized residuals for Moderate Left, Communist and

Other parties

Moderate Left

+————————————-+

| département year stud.resid |

|————————————-|

41. | Alpes-Maritimes 2002 -2.19866 |

261. | Jura 1986 2.316615 |

293. | Haute-Loire 2002 -2.025205 |

352. | Haute-Marne 1986 2.060687 |

355. | Haute-Marne 1997 -2.564131 |

|————————————-|

408. | Oise 1986 2.182928 |

426. | Pas-de-Calais 2002 -2.22702 |

464. | Haut-Rhin 1986 2.41543 |

534. | Yvelines 1986 2.006661 |

573. | Var 2002 -2.169649 |

|————————————-|

659. | Corse-du-Sud 1981 3.205233 |

660. | Corse-du-Sud 1986 4.251734 |

663. | Corse-du-Sud 1997 -2.248766 |

664. | Corse-du-Sud 2002 -4.478615 |

+————————————-+

Communists

+———————————————+

| département year stud.resid |

|———————————————|

13. | Aisne 2002 -3.365818 |

23. | Alpes-de-Haute-Provence 1986 2.262079 |

27. | Alpes-de-Haute-Provence 2002 -2.581943 |

50. | Ardennes 1981 2.240427 |

55. | Ardennes 2002 -4.792705 |

|———————————————|

76. | Aude 2002 -2.099659 |

111. | Charente 2002 -2.66234 |

181. | Eure 2002 -2.27876 |

202. | Gard 2002 -2.175194 |

230. | Hérault 2002 -2.377525 |

|———————————————|

377. | Meuse 2002 -2.454502 |

426. | Pas-de-Calais 2002 -2.64686 |

573. | Var 2002 -3.137405 |

580. | Vaucluse 2002 -2.145657 |

597. | Haute-Vienne 1986 2.517245 |

|———————————————|

608. | Vosges 2002 -2.402605 |

+———————————————+

Other

+---------------------------------+
| département year stud.resid |
|---------------------------------|
43. | Ardèche 1981 -4.204402 |
52. | Ardennes 1988 -4.296376 |
73. | Aude 1988 -3.726954 |
80. | Aveyron 1988 -4.251042 |
82. | Aveyron 1997 2.003402 |
|---------------------------------|
108. | Charente 1988 -3.450763 |
122. | Cher 1988 -3.783795 |
129. | Corrèze 1988 -3.623167 |
148. | Creuse 1981 -3.935467 |
157. | Dordogne 1988 -3.538015 |
|---------------------------------|
178. | Eure 1988 -3.736878 |
239. | Indre 1981 -4.06781 |
269. | Landes 1988 -3.845304 |
288. | Haute-Loire 1981 -3.965096 |
309. | Lot 1981 -4.078404 |
|---------------------------------|
360. | Mayenne 1988 -3.709984 |
374. | Meuse 1988 -3.643376 |
416. | Orne 1988 -3.594169 |
465. | Haut-Rhin 1988 -3.40009 |
479. | Haute-Saône 1988 -4.197225 |
|---------------------------------|
500. | Savoie 1988 -3.453629 |
540. | Deux-Sèvres 1981 -4.398553 |
556. | Tarn 1988 -4.052777 |
584. | Vendée 1988 -3.667911 |
605. | Vosges 1988 -3.662599 |
|---------------------------------|
666. | Haute-Corse 1981 -4.380981 |
+---------------------------------+

1 Replication data are available from the authors’ dataverse at XXXXX

1Notes

Table 1 Nested model comparisons, fixed effects and full model

forecasts (1981-2002)

Model

ll(null)

ll(model)

df

R2

BIC

Moderate Left: Fixed Effects

-394.22

-248.17

96

0.40

1106.54

Moderate Left: Full Model

-121.87

100

0.61

879.36

Moderate Right: Fixed Effects

-244.49

-135.97

96

0.31

882.12

Moderate Right: Full Model

-301.37

100

0.57

641.64

Communists: Fixed Effects

-633.09

-439.15

96

0.49

1487.98

Communists: Full Model

-374.60

100

0.59

1384.30

Extreme Right: Fixed Effects

-889.66

-850.12

96

0.14

2301.53

Extreme Right: Full Model

-643.54

100

0.61

1913.42

Other: Fixed Effects

-1445.50

-1414.72

96

0.11

3435.89

Other: Full Model

-1320.19

100

0.36

3272.09

Table 2 Parameter estimates for full model forecasts (1981-2002)

Comm

ML

MR

ER

Other

Unemployment (U)

-0.03 (.01)

-0.10 (.01)

-0.06 (.01)

0.53 (.03)

0.63 (.09)

Left incumbency (I)

-1.69 (.10)

-0.66 (.12)

-0.36 (.10)

3.14 (.35)

5.24 (1.04)

ΔGDP

0.00 (.01)

0.12 (.01)

0.05 (.01)

0.29 (.03)

-0.79 (.09)

U*I

0.14 (.01)

0.08 (.01)

0.08 (.01)

-0.15 (.03)

-0.46 (.10)

R2

.59

.61

.57

.61

.36

Root MSE

.51

.33

.27

.92

2.90

– All contrasts with abstention baseline

– Departmental dummies not shown

– All coefficients significant at .05 or lower, except Comm ΔGDP

Table 3 Predicted and observed vote outcomes in 2007 (% of total

registered voters and of votes cast)

% registered voters

% votes cast

Observed

Predicted

Observed

Predicted

Communists

2.7

8.8

4.4

13.3

Moderate Left

16.9

26.6

27.7

40.1

Moderate Right

32.6

29.0

53.4

43.7

Extreme Right

2.9

1.2

4.8

1.8

Other

5.8

0.9

9.5

1.4

Abstention

39.0