Electoral volatility in Belgium (2009–2014). Is there a difference between stable and volatile voters?

Abstract

Increasing voter volatility has led to a renewed research interest in determinants of party switching. While previous research has mainly focused on the characteristics of volatile voters, less is known about how stable and volatile voters decide what party to vote for. Using panel data spanning two consecutive electoral cycles in Belgium, this study starts with the confirmation of earlier findings: we show that widely used determinants like political sophistication and disaffection add only modestly to our understanding of volatility. In a next step, we examine the vote choice process of stable and volatile voters. Our results indicate that in terms of determinants of the vote choice the two groups are somewhat different. In line with theoretical expectations about the effects of stronger voter volatility, we find that party-switchers are guided more by proximity evaluations. The implication of these results is that party-switchers might actually be enriching representative democracy. We close with some observations on how this finding qualifies our theoretical understanding of increasing levels of electoral volatility in liberal democracies.

Appendices

Appendix A: Measurement and Coding of the Variables in the Analyses Explaining the Characteristics of Stable and Volatile Voters

Sex: sex of the respondent: 0 = male, 1 = female.

Age: age of the respondent, calculated by subtracting the year of birth from the year in which the interview is conducted.

Region: region in which the respondent lives: 0 = Dutch-speaking region (Flanders), 1 = French-speaking region (Wallonia).

Religious practice: answer on the question how often the respondent participates in religious activities. Scale ranging from 0 (never) to 6 (once a week or more often).

Social class: social class of the respondent, divided in three categories: inactive (retired, student, unemployed, houseman/wife, other not active), self-employed, manual worker and non-manual worker.

Political trust: sum scale of questions probing trust in political parties; the regional government, the regional parliament, the federal government and politicians. For each institution, the respondent could indicate her/his trust on a scale ranging from 0 (no trust at all) to 10 (fully trust). The answers on these questions load on one factor ‘political trust’ (eigenvalue = 3.908; Cronbach’s Alpha = 0.94).

Government satisfaction: answer to the question how satisfied the respondent is with the policy pursued by the regional government on a scale from 0 (very unsatisfied) to 4 (very satisfied). This question was probed with respect to the government of the region the respondent lives in.

Political knowledge: sum scale of answers on five knowledge questions. For each question, the respondent was given four possible answers to choose from. These have been recoded to indicate a correct (1) or wrong or ‘don’t know’ (0) answer, and subsequently added up to construct one scale.

Political interest: answer on the question how interested the respondent is in politics in general. Respondents could indicate their interest on a scale ranging from 0 (not at all interested) to 10 (very much interested).

Educational level: the level of education of the respondent. Respondents are divided in three categories: (0) no degree/primary degree/uncompleted secondary degree, (1) completed secondary degree, (2) higher education/university (Table A. 1).

Table A.1 Descriptive statistics of variables in the logit analyses

Appendix B: Measurement and Coding of the Variables in the Analyses Explaining the Vote Choice Process of Stable and Volatile Voters

Party identification: indication of the extent to which the respondent identifies with a party, ranging from 0 (no party identification) to 4 (very close).

Distance to party: distance between the respondent’s self-placement on the ideological axis ranging from 0 (left) to 10 (right) and the party’s position on this axis. The parties’ positions were calculated by taking the average of the positions the respondents assigned to the different parties.

Leader evaluation: respondent’s evaluation of the leader of the parties, respectively, on a scale from 0 (no appreciation) to 10 (a lot appreciation).

Economic evaluation: respondent’s retrospective evaluation of the Belgian economic situation on a scale from −2 (strongly decreased) to 2 (strongly increased).

Party in government: indication of whether the party was in government before the elections of 2014 (code 1) or not (code 0) (Tables B.1, B.2).

Table B.1 Descriptive statistics of variables in the fixed effects conditional logit analyses: stable voters

Table B.2 Descriptive statistics of variables in the fixed effects conditional logit analyses: volatile voters

Appendix C: Logit Analyses Explaining Electoral Volatility

See Tables C.1, C.2 and C.3

Table C.1 Logistic regression models explaining (individual-level) electoral volatility 2009–2014 with political interest as indicator of political sophistication

Table C.2 Logistic regression models explaining (individual-level) electoral volatility 2009–2014 with educational level as indicator of political sophistication

Table C.3 Logistic regression models explaining (individual-level) electoral volatility 2009–2014 with political knowledge as indicator of political sophistication

Appendix D: Average Marginal Effects of the Models Displayed in Table 1. These Average Marginal Effects are Displayed in Figure 2

See Table D.1.

Table D.1 Average marginal effects of the models displayed in Table 1

Appendix E: Analyses of the vote choice process for the Belgian regions separately

See Tables E.1 and E.2.

Table E.1 Fixed effects conditional logit models explaining the vote choice process of stable and volatile voters in the regions separately

Table E.2 Average marginal effects of mixed effects conditional logit models explaining the vote choice process of stable and volatile voters in the regions separately (Table E.1.)

Appendix F: Weighting the Analyses

In order to provide a solution for panel attrition in the data, we calculated attrition weights. The method that we used is ‘Inverse Probability Treatment Weighting (IPTW) Using the Propensity Score – the same method that has been used to weigh the GLES data (Blumenstiel and Gummer, 2013). This method creates a weighting variable in which the representativeness of the panel respondents is ensured (Austin, 2011). Panel data are particularly suited for the IPTW method since by definition in the first wave of the study all respondents (so also future non-respondents) took part and completed the questionnaire. In contrast to what holds for cross-sectional studies, thus, there are data available on the characteristics of non-respondents from the second wave on (Rizzo, Kalton and Brick, 1994). This information of the first wave makes it possible to calculate weights specifically created to correct for panel attrition (Vandecasteele and Debels, 2007).

The rationale behind IPTW is the following. In a first step, the determinants of dropout are searched for. Subsequently, based on this knowledge, the probability for each respondent to participate in the next wave is calculated. The attributed weight, then, is the inverse of this probability. By taking into account the appropriate determinants, this inverse of the probability of participating thus ensures that those respondents who have the profile of respondents of whom we would expect to drop out but actually do participate receive a higher weight than those of whom it could be expected that they would continue participation. Thus, in this way, the results are corrected to the factors driving dropout/participation, and ensure a more equal representation.

Since the IPTW technique is based on predicting – for every respondent – the probability of participating in a next wave by means of observed information in the previous wave, it is important to consider well which variables to take into account in these regressions. The covariates that should be included are the variables that are related to panel (non-)participation. These covariates can be traced by including them in a series of regressions of which the explanatory power is compared. Using this method, the variables that add to the explanation of participation in the next wave can be included, and the variables not adding any explanatory power can be dropped. To make an appropriate weight for the BEP data, we start with a general model including variables of the first wave to explain the participation in the second wave. The variables that are included in this general model have been selected on the basis of earlier research on panel attrition (Olson and Witt, 2011). Next, we searched for other variables that help explain continued engagement in the study. We do this by estimating series of logistic regression models of which we compare the explanatory power, and keeping the variables that increase this power with more than 0.05%

Proceeding in this way, we calculated the attrition weights based on a model that predicts participation in subsequent survey waves best. The determinants included in our final model were region of residence (i.e. Flanders or Wallonia), sex, age, occupation, political knowledge, willingness to turn out, vote intention, political interest, political trust, marital status, number of children, political efficacy, newspaper reading, listening to radio, visiting a party website, visiting website politician, reading press of party, discussion with colleagues and discussion with family.

A logistic model predicting the probability of participation in the later waves including these variables is thus estimated, and every respondent is then assigned the weight that corresponds with the inverse of this probability. In Table F.1, we display the results of the model explaining participation in the fifth wave, on the basis of which we calculated the attrition weight used to weigh the analyses in this paper.

Table F.1 Logistic regression model explaining participation in wave 5 of the Belgian Election Panel

It is important to note that the weight for the fourth and fifth waves of the study contain rather high values. In line with what is commonly done in such a case (Chen et al, 2015), we therefore winsorise the weights setting as a maximum value three times the median value in the sample. However, this reduces the total weighting since the largest values are fixed to a lower value. We correct for this by calculating the total weight before and after the winsorising. Then, we multiply each weight with the ratio of these two, ensuring the total weighting after winsorising equals the total weighting before winsorising.

This method does not allow providing a correction for non-response in the first wave of the panel study but serves to account for panel attrition in subsequent panel waves. As an alternative, the data from the first wave are weighted to the known population distributions using the method of Iterative Proportional Fitting. With this weight, the representativeness of the first wave is ensured. Since the IPF-weight ensures the representativeness of the general dataset, combining this weight with the specific weights correcting for panel attrition combines the two correcting mechanisms. Therefore, the panel-weights that account for attrition in the data are multiplied with the IPF-weight that ensures the representativeness of the first wave, which gives us the final weights. Note that these final weights were winsorised as well using the procedure described above.

Appendix G: Analyses Using Only Party Minister-President as ‘Party in Government’

See Tables G.1 and G.2.

Table G.1 Fixed effects conditional logit models explaining (individual-level) electoral volatility 2009–2014 with only the party of the Minister-President as ‘party in government’

Table G.2 Average marginal effects of vote choice determinants explaining voting in the 2014 regional elections