Forecasting the rise of the Front National during the 2014 municipal elections

Data, Measures and Methods Forecasting the rise of the Front National during the 2014 municipal elections Sylvain Brouard and Martial Foucault* CEVIPOF, CNRS, Sciences Po, 98 rue de l Université, 75007 Paris, France. E-mails: sylvain.brouard@sciencespo.fr; martial.foucault@sciencespo.fr *Corresponding author. Abstract This article develops an electoral forecasting model for the extreme-right vote share in France during the 2014 municipal elections. On the basis of data gathered between 1998 and 2014 from a sample of 56 cities where the Front National (FN) has always presented candidates for municipal office, the model anticipated a rise in the FN s share of the vote during the 2014 municipal elections. Controlling for political context through election type, FN popularity and electoral dynamics in addition to the criminality rate, this article adds to previous extreme-right forecasting models demonstrating that the FN s vote share is not at all unpredictable even in the case of local elections. French Politics (2014) 12, 338 347. doi:10.1057/fp.2014.19 Keywords: extreme-right; electoral forecasting; municipal elections Introduction Despite low levels of media coverage and scant mobilization among the political parties, the 2014 French municipal elections were a tipping point for Hollande s presidency, accelerating a change of prime minister, a government reshuffle including an albeit limited number of new ministers and the replacement of the head of the governing party (the Parti Socialiste). Such changes were mainly because of the huge defeat suffered by left-wing incumbent candidates in cities where they had governed since the previous municipal elections in 2008. Furthermore, this blue wave, as the rise of the French extreme-right party the Front National (FN) is known, has cemented the position of third place gained by its leader, Marine Le Pen, during the 2012 presidential election. Somewhat surprisingly, the FN has never been structured to ensure its candidates get elected at municipal level. France is divided into 36 560 municipalities. Roughly 1080 cities within those municipalities have over 9000 inhabitants leading to a more politicized ballot. As a result, every party www.palgrave-journals.com/fp/

Forecasting the rise of the Front National must convince thousands of candidates to register on electoral lists before the election. For small parties, municipal elections represent a major challenge in terms of mobilizing candidates and structuring their campaigns. The large increase in FN candidacies may well have affected the outcome for the traditional left and right parties in these elections, where the issues are not only local but also national. The FN obtained its best electoral outcome in 1995 obtaining 13.1 per cent of the votes in the larger municipalities where it presented candidates, a score that enabled it to win over 10 per cent of valid votes in more than 290 cities and to have almost 1000 municipal councillors elected. In 2014, the question is to ascertain to what extent the FN is likely to perform as well as in 1995. In this article, we present a methodological approach used to forecast the FN s electoral outcome during the 2014 municipal elections. The model, which was published in March 2014 (See www.slowpolitix.blogspot.fr/2014/03/municipales- 2014-une-estimation-du.html; www.500signatures.net/index.php?id=32), measures the FN s share of the first round vote in cities where the party consistently presented lists between 1995 and 2014. In such a way, we investigate a set of explanatory variables identifying the political strength of the FN both at the local and national level. On the basis of a panel data estimator with fixed effects, we provide accurate forecasts from 2001, although highly accurate forecasts at the city level were harder to obtain. Contrary to poll surveys, this model performs much better in anticipating the rise or decline of the FN in local elections. The FN and Municipal Elections As part of Marine Le Pen s strategy from 2011, the FN has made great efforts to cover all French electoral constituencies by encouraging candidates to run for office. As the party suffers from a lack of representation, Le Pen has targeted local elections as the main channel to build up a representative basis with the 2017 presidential election in mind. As a consequence, about 600 cities were identified as municipalities where the FN could get municipal councillors elected based on the party s performance during the 2012 presidential election. On 6 March 2014, 3 weeks before the first round of the municipal elections, the number of electoral lists presented or supported by the FN was officially submitted to the prefectures in charge of organizing municipal elections. In municipalities of more than 1000 inhabitants the FN presented 596 lists more than ever before. With nearly 450 lists presented in cities of over 9000 inhabitants, the FN aimed at renewing its organizational capacity of 1995 despite the new constraint of parity regarding the composition of the lists. 1 For the record, the FN presented lists in 303 municipalities of more than 9000 inhabitants in 1989, 456 in 1995, but only in 183 and 106 cities, respectively, in 2001 and 2008 (to which can be added, respectively, 114 and 5 cities 339

Brouard and Foucault in which no FN candidate ran for election but where extreme-right lists were presented). Despite the increasing number of municipalities where the FN fielded candidates, only a few of them (51 cities) have had consistently registered FN (or extreme-right) candidates since 1995. They include some very small cities such as Harnes, Noyon, Lunel, Vernon and so on, and larger cities such as Marseille, Lyon, Paris and Strasbourg. Such heterogeneity makes it both difficult and costly to obtain accurate surveys. Moreover, as interesting as they may be, it is very difficult for opinion polls on voting intentions to forecast realistic scores for the FN. This is because many FN voters tend not to declare a voting intention for the party. Added to this, such surveys only report the estimated vote in a dozen cities, which are not always representative of the national trend as a whole, and which therefore do not permit the FN s overall score to be predicted. In order to overcome such difficulties, an alternative method was successfully tested. The method was based on the modelling of former election results to predict the national scores of the FN since 1984 (Evans and Ivaldi, 2010) and of Marine Le Pen in the 2012 presidential election (Evans and Ivaldi, 2012). Today, such a perspective has not been implemented either at the disaggregated level of the cities or for FN votes in municipal elections. We therefore built a simple forecasting model to identify the explanatory variables of the FN s vote at the local level since 1995 (with the exception of legislative elections) and then estimated the FN s score in the first round of the municipal elections in March 2014. This methodological choice deserves some further explications. Model The basic principle according to which an FN forecasting model is built depends on what function individual voters attribute to their vote. The vote cast will reflects a vote function considering a combination of local and national factors. For example, even if an incumbent mayor is accountable for her past performance, her notoriety and her policy promises, voters may nevertheless use the opportunity presented by the election to reward or punish the national government for its performance as well as the opposition parties at the national level. From this perspective, it is therefore essential to consider the FN s popularity at the national level as a proxy of sympathy for tier parties. Further, the electoral dynamics of every city studied is a determinant variable for the simulation of the FN s future score. In Paris, the party s average score was 6.6 per cent between 1998 and 2012, whereas it was about 25.5 per cent in Carpentras in the south of France during the same period. The FN s volatile results from 1995 to 2008 (see Table 1) at the municipal level become part of a much more complex dynamic structure when considering also its results in national, regional and European 340

Forecasting the rise of the Front National Table 1: Panel data estimates, 1995 2012 (DV = FN vote share) Within forecasts Step-ahead forecasts Before 2014 After 2014 2001 2008 2014 FN vote share (t-1) 0.121*** 0.0828** 0.213*** 0.196*** 0.121*** (0.0349) (0.0328) (0.0542) (0.0604) (0.0349) FN vote share _national 0.545*** 0.522*** 0.812*** 0.146** 0.545*** (0.0265) (0.035) (0.0665) (0.0556) (0.0265) FN popularity 0.437*** 0.335*** 1.76*** 0.82*** 0.437*** (0.0407) (0.04) (0.141) (0.149) (0.0407) Cohesion 6.04*** 5.84*** 7.75*** 6.04*** (0.545) (0.629) (0.604) (0.545) Criminality rate 2.07*** 2.02*** 1.23*** 2.07*** (0.0855) (0.0831) (0.111) (0.0855) Regional 7.22*** 7.15*** 5.21*** 7.22*** (0.48) (0.464) (0.736) (0.48) Municipal 1.28** 0.36 0.869 1.28** (0.508) (0.621) (0.748) (0.508) Constant 0.307 1.92** 14.5*** 20.3*** 0.307 (0.586) (0.856) (2.14) (2.13) (0.586) Observations 616 666 112 392 616 Adjusted R 2 0.716 0.669 0.961 0.719 0.716 Akaike Information Criterion 3147 3528 312 1945 3147 Sum of Square Errors 3.76 4.11 3.48 3.71 3.76 *P<0.1, **P<0.05, ***P<0.01. Standard errors in parentheses. Note: City fixed effects are not reported because of space limitation. elections in each city. For instance, in Dreux where the FN attained an average score of 14 per cent of the votes between 1998 and 2012, the scores ranged from 4.3 per cent in the 2008 municipal election to 31.7 per cent in the 1998 regional election. Insofar as our aim was to forecast the FN vote (VOTESHARE t as a dependent variable) during the 2014 municipal elections, we selected 56 cities in which the FN and/or former party members presented electoral lists in the municipal elections of 1995, 2001 and 2008. On the basis of the scores of the FN in the municipal, presidential, regional and European elections from 1998 to 2012 (as a dependent variable), our statistical model aims at identifying the relative weight of diverse predictive variables of the FN s vote. Contrary to previous models of electoral forecasting, we offset the small number of municipal cases (56 cities 2 municipal elections) by adding the FN vote share for European, presidential and regional elections at the city level. By doing so, we captured the real dynamics of the FN vote over time and not only for municipal elections, which only occur every 6 years in France. 341

Brouard and Foucault Among independent variables, we made a trade-off between relevant variables, lead variables and a parsimonious model (Lewis-Beck, 2005). Consequently we paid great attention to electoral trends (the FN vote share during the preceding municipal elections, VOTESHARE t-1 and the FN vote share during the preceding national election, VOTESHARE_NAT t ), the FN s popularity (POP t ), the rate of criminality (CRIME t ), dummies to account for the cohesion of the party (COHESION t ) and local elections effects (REGIONAL t and MUNICIPAL t ). As we were not able to capture all city effects, we added a fixed-effect specification (λ i ) and clustered standard errors at the city level to correct heteroscedasticity. The forecasting model has the following form: VOTESHARE i;t ¼a + β 1 VOTESHARE i;t - 1 + β 2 VOTESHARE NAT i;t - 1 + β 3 POP t + β 4 CRIME t + β 5 COHESION t + β 6 REGIONAL i;t + β 7 MUNICIPAL i;t + λ i + ε i;t Our expectations about the sign of each coefficient are derived from the following hypotheses: The higher the FN s score was in the preceding municipal election and in the preceding national election, the higher their score would be in the subsequent election (β 1 >0; β 2 >0). The more the FN s popularity grows, the more their score would increase (β 3 >0). The more the crime rate rises, the more their vote share would increase (β 4 >0). When the FN is divided, its score declines (β 5 <0). The sign for REGIONAL and MUNCIPAL is not directly derived from a theoretical hypothesis. Those variables act as controls to capture the heterogeneous ability of extreme-right candidate to perform in local elections because of the lack of candidacies and local resources. As our unit of observation is the city in which the FN has always presented candidates, we suspect there is a positive relationship between these variables and vote share. Results Panel data was used to estimate the model (t = 11 elections; i = 56 cities; N = t i = 616). 2 The political and criminality information included in the specification cannot be assumed to account for all possible cofactors that might influence the vote, such as religion, culture, socio-demographic indicators and so on. The solution to this problem is to use a fixed-effect model to control for unobserved time-invariant differences between cities (practically speaking, this means including i 1 city dummy variables in the model; see λ i parameter in the equation above). Using this 342

Forecasting the rise of the Front National estimation technique ensures that the estimated coefficients presented in Table 1 are not biased because of omitted time-invariant variables. Another potential problem could arise because both popularity and criminality variables (POP) are not only time-variant but also space-invariant (that is, their value is the same for all cities in a given year). To ensure that our estimates are not biased by a time random effect (see Greene, 2008), we performed Hausman tests (with and without the variables POP and CRIME in the model). They confirm that the city-fixed effect and time-random effect model is the adequate specification for our data. On the basis of 616 observations, our model reports statistically significant coefficients for each embedded independent variable. It explains a large proportion of the FN s score variance (R 2 = 0.7) with a standard error estimate rounded to 3.76. The general specification offers a within or in-sample estimation from which we forecast the FN vote share for 2014. As expected, all coefficients are well-signed and significantly different from zero. Basically, this model does not differ greatly from the classical vote function elaborated by Kramer (1971) and assessed by Nannestad and Paldam (1994) with an emphasis on political and issues contexts. Contrary to Jérôme and Jérôme-Speziari (2003) and Evans and Ivaldi (2008), we do not include an economic variable, such as unemployment. The reason is twofold: first, there is no available data at the city level (specially for small cities) over the period under investigation; second, the relationship between unemployment and the extreme-right in France is not totally exogenous as pointed out by Martin (1996), which implies that a valid instrument must be found to obtain unbiased covariates. The flow of immigrants in a city or neighbouring cities (the halo effect) might well provide a future avenue for research. An alternative would consist of including the national unemployment rate as a common shock for all cities, which is what we did. Nevertheless, whatever the specification (level or annual/quarterly rate of change), we did not find a significant relationship between unemployment and the FN s vote share. Clearly, any forecasting exercise implies parsimony. Thus, adding a nonsignificant variable increases the degrees of freedom without improving the accuracy of the out-of-sample forecasts. Before turning to the forecasting, two methodological precisions should be made. A forecasting model relies on a parsimonious and robust statistical model, which has to be replicable whenever used, based on the explanatory variables known before the election (lead). Our model is based solely on variables known before the election, such as, for example, the popularity of the FN, measured by the TNS Sofres Figaro Magazine barometer (www.tns-sofres.com/dataviz?type=3&code_nom=fn), respectively, held in December of the year before the municipal, presidential and regional election as well as in March before the European elections. Second, in Table 1 we present the step-ahead forecasts to better understand how our model evolves over time and is completed by the addition of new observations. Interestingly, in 2001 the model reports a high R 2 with only three independent variables because of the lack of data for preceding years. However, the unexpected negative sign for 343

Brouard and Foucault VOTESHARE_NAT indicates that the FN did not manage to take advantage of the 1999 European parliament election because of the fragmentation of the party between supporters of the party leader Jean-Marie Le Pen and partisans of Bruno Mégret, who challenged his leadership at that time. In 2008 and 2014, the model comes closer to the general specification used to forecast out-of-sample estimates (Column 1). Table 2 reports the quality of forecasts for all elections considered. The average FN score for the different elections in the 56 cities is close to the actual electoral outcome, with some significant discrepancies for the presidential elections in 2007 ( 4.11 per cent) and in 2012 (3.36 per cent). In close to three-quarters of cases, the difference between the predicted and the actual FN vote share is less than 2 points. The outcome of the 2004 regional election was most accurately predicted, with the model missing the mark by only 0.12 points. On average, the model reports an absolute error of 1.84 points that must be put into perspective with the recurrent underestimation of extreme-right parties by pollsters in France. Regarding municipal elections, our model performed quite well in 2001 and 2008, and a little less well in 2014. On the basis of the variables known in December 2013, the model forecast an average score of 22.24 per cent for the FN in the 51 cities 3 under investigation with a standard error estimate of 3.76. In reality, the FN reached 19.67 per cent in March 2014, that is, 2.5 times higher than in 2008. Such a performance is in line with our forecasts even if the model tends to overestimate the actual outcome after the first round. Clearly, compared with 2008, the FN has benefitted from Marine Le Pen s modernization of the party. Nevertheless, the same model performed better in 2001 and 2008 (see Table 3) although it underestimated the real vote share. Two factors played a role in this change. First, the FN s popularity rate has risen substantially in Table 2: Performance of the out-of-sample model Type of election Year Forecasts Actual score Difference Regional 1998 21.3 22.53 1.23 European 1999 14.99 13.52 1.47 Municipal 2001 12.77 14.21 1.44 Presidential 2002 21.1 23.12 2.02 Regional 2004 20.79 20.66 0.12 European 2004 11.72 13.1 1.38 Presidential 2007 15.84 11.73 4.11 Municipal 2008 6.54 8.14 1.6 European 2009 10.74 9.47 1.27 Regional 2010 17.88 16.16 1.72 Presidential 2012 15.62 18.98 3.36 Municipal 2014 22.24 19.67 2.57 Average 1.85 Source: Data from the Ministry of the Interior, EDEN (CEVIPOF), authors. Bold values signify municipal elections. 344

Forecasting the rise of the Front National Table 3: Out-of-sample and within forecasts, 1998 2014 (in percentage) 2014 2008 2001 Forecast (within) 20.63 7.65 13.81 Actual 19.67 8.14 14.21 Mean absolute error 0.96 0.51 0.60 Largest absolute error 21.35 17.00 18.50 Forecast (out-of-sample) 22.24 6.54 12.77 Actual 19.67 8.14 14.21 Mean absolute error 2.57 2.40 1.56 Largest absolute error 19.46 18.15 20.25 recent years, increasing from 8 per cent in March 2008 to 23 per cent in December 2013. Consequently, the value for the out-of-sample predicted variable has made significant gains as a result of the increase in favourable opinions of the FN subsequent to Marine Le Pen s arrival as head of the party. Second, with a 0.54 coefficient, the last national electoral outcome (the 2012 presidential election) provides a net advantage for cities where Marine Le Pen performed well in 2012 (16.1 per cent on average). Discussion To some extent our model was as accurate in 2014 as it was in 2008 with a mean absolute error close to 2.5 (see Table 3). Interestingly, our forecasts predicted the rise of the FN in the 2014 municipal election well before it happened. Even if our sample contains only cities where the FN has fielded candidates since 1995, no bias selection has affected our forecasts. Table 3 compares the out-of-sample forecasts (one election omitted) with within forecasts (all years included) and reports as expected a better fit for the within specification (mean error at 0.96). However, in some cities (such as Hénin-Beaumont or Béziers), the model fails to capture the FN s high score. As the model was not designed to forecast at the city level but rather at the aggregate level, this means that such discrepancies could be corrected in the future by taking advantage of local measures of popularity or criminality. At the same time, the addition of new variables (such as immigration flows, accountability measures for the incumbent) could be explored and a trade-off made between adding explanatory variables and maintaining the parsimony principle of any forecasting specification. Our empirical strategy aimed at better calibrating a model combining various elections and using for the first time city as unit of observation. Forecasting an election result for the French extreme-right at the municipal level is a tough challenge 345

Brouard and Foucault Figure 1: Residuals versus fitted values. as the number of observations is low (56 cases per election) and out-ofsample residuals do not roughly form a horizontal band around the 0 line (see Figure 1). This suggests that error terms are not perfectly equal when looking at specific election years but fit better when the period as a whole is studied (Figure 1 bottom right). Our estimation for the first round of the 2014 municipal elections predicted a historical surge for the FN, particularly in comparison to other municipal elections. The model forecast that the FN would obtain 22.24 per cent of the vote, which is an overestimation of the party s actual outcome in a sample of 51 cities by 2.57 points. Even if the Holy Grail of any forecasting exercise is to converge towards a null error, our model remains below the standard error estimate (3.76 points) and confirms that the selected independent variables are good predictors of the FN s vote share. Replicating the model for other elections and for other types of constituency would be a relevant test of its robustness. Notes 1 On 6 June 2000, France passed a gender parity law stating that all political parties should include equal numbers of men and women on party lists for cities of more than 3500 inhabitants. This threshold was 346

Forecasting the rise of the Front National modified in 2013 to include all cities of over 1000 inhabitants. Nevertheless, even though the FN respected the law, only 18 per cent of women were front runners on their lists in March 2014. 2 Municipal elections in 2001, 2008; presidential elections in 2002, 2007 and 2012; regional elections in 1998, 2004 and 2010; and European elections in 1999, 2004 and 2009. 3 In five cities (Saint-Louis, Enghien-les-Bains, Romainville, Neuilly sur Seine and Bourg-de-Peage) no FN candidate ran for office in 2014. References Evans, J. and Ivaldi, G. (2008) Forecasting the extreme-right vote in France (1984 2007). French Politics 6(2): 137 151. Evans, J. and Ivaldi, G. (2010) Comparing forecast models of radical right voting in four European countries (1973 2008). International Journal of Forecasting 26(1): 82 97. Evans, J. and Ivaldi, G. (2012) Forecasting the extreme-right vote at the 2012 presidential election: Evaluating our model. French Politics 10(4): 378 382. Greene, W.H. (2008) Econometric Analysis, 6th edn. Upper Saddle River, NJ: Prentice Hall. Jérôme, B. and Jérôme-Speziari, V. (2003) A Le Pen vote function for the 2002 presidential election: A way to reduce uncertainty. French Politics 1(2): 247 251. Kramer, G. (1971) Short-term fluctuations in US voting behaviour, 1896 1964. American Political Science Review 65(1): 131 143. Lewis-Beck, M.S. (2005) Election forecasting: Principles and practice. British Journal of Politics and International Relations 7(2): 145 164. Martin, P. (1996) Le vote le Pen: l'électorat du Front national. Paris, France: Note de la Fondation Saint- Simon. Nannestad, P. and Paldam, M. (1994) The Vp-function A survey of the literature on vote and popularity functions after 25 years. Public Choice 79(3 4): 213 45. 347