association between the general wealth of a country and the polarisation of its party system. This association, thus, remains ambiguous.
There is, however, an association, which has been rather explicitly stated and accepted, but which has not been empirically tested at all – one with the level of democracy. This is something that was clearly present in the works of Downs (1957) and Sartori (2005), but which has not made its way into the analysis of polarisation yet. In brief, the argument of Downs was that ideological polarisation will make government policy across governments unstable, which can result in political chaos and possibly even revolution (Downs 1957, p. 120). The argument of Sartori was similar – anti-system parties at the extremes result in a politics of outbidding, which undermines the political life of a country. If the extremes become too strong and overtake the stabilising centre, the political system can fall apart. Therefore, one could expect an association between increased polarisation and lower levels of democracy. With democracy measures abound (e.g. Coppedge et al. 2016b), this is rather straightforward to test empirically.
(see Section 5.1.3).
The data that is used in the comparison below is obtained for the most part form the 2014 stable version of the ParlGov data set (D¨oring and Manow 2014), the Standardized World Income Inequality Database (Solt 2016) (based on data from the Luxembourg Income Study), the Quality of Governance data set (Teorell et al. 2016) and the Varieties of Democracy data set (Coppedge et al.
2016b). The objective of this analysis is not to test specific associations with polarisation (although this will also be comment on below), but to provide a plausible model given what we know about the possible associations with polarisation in order to compare the measures for party differences.
The analysis will therefore focus only on those variables that have been central to the theories and analyses that were discussed above and will not focus on the myriad of “control” variables that some of the analyses have included. Thus, the model includes the following (the abbreviations in capital letters are used in the figures and tables below):
• Polarisation (POL). Based on the following measures of party position or difference (see Section 3.3):
– EELR Elff’s left-right scale
– FKLRFranzmann and Kaiser’s left-right dimension – JJahn’s left-right dimension
– KK¨onig et al.’s left-right dimension
– KFRILEversion of RILE proposed by Kim and Fording – LRILERILE using the logit scale of Lowe et al.
– PLRProsser’s left-right dimension
– RILE left-right index of the manifesto data set – SIMthe index of similarity
Using these indices, polarisation is operationalised through the weighted average pairwise dis-tance measure (PW) (equation 5.7) and the ideological standard deviation (SD) (equation 5.2) measure. The values are standardized for the final set of cases so that the model coefficients would be comparable in their substantive magnitude.
• Fragmentation (FRAG).Measured as the effective number of parliamentary parties (Laakso and Taagepera 1979), calculated from the ParlGov data.
CEUeTDCollection
• Disproportionality. Measured as the Gallagher index (Gallagher 1991, p. 40), which is based on the difference between seat shares and vote shares, calculated from the ParlGov data set.
• Voter turnout. Percentage of the electorate that turned out to vote. Obtained from the Quality of Governance data set.
• Electoral volatility. Volatility is separated into that, which happens among established (con-tinuous) parties and that, which can be attributed to the emergence of new parties. The data on volatility is obtained from the Dataset of Electoral Volatility and Its Internal Components in Western Europe (1945-2015) (Emanuele 2015).
• Government stability. The average duration of governments (in months) that were formed during the parliamentary term that is associated with a given election. Calculated from the ParlGov data set.
• Level of democracy. Measured using the electoral democracy (polyarchy) index form the Varieties of Democracy data set. This index is supposed to reflect the most basic aspects of democracy – a democratically functioning electoral process.
• Inequality. Household inequality before taxes, Gini index. Obtained from the Standardized World Income Inequality Database (Solt 2016), which in turn uses data from the Luxembourg Income Study. The original data set contains 100 estimated values for each included country year. Their mean as the best estimate for any given year is used in this analysis. The data set does not cover every year and thus it can happen that for a given election year, data is missing. In such cases data from up to three years in the past is used.5
• GDP growth and inflation. In some of the analyses that were mentioned above these are just used as “control” variables, but they reflect important indicators of economic conditions, which can influence party positions and are thus included here. Data for both is obtained from the Variates of Democracy data set.
• Coalition habits. Instead of a dummy variable indicating whether coalitions or minority governments formed in the past, the current comparison uses a more nuanced measure that should capture the same underlying phenomenon better. One would need a measure that captures not only if coalitions form, but also how open they are. If there are two fixed party blocs
5 If there were still gaps in the middle of a continuous series of data, they were filled with averages of the adjacent values. Of all the data that is used here, data on inequality is most restricted in terms of availability as for most countries it is available only from the 1970s or 1980s onwards. Losing further cases would thus undermine the analysis more than using adjacent values for the missing election years.
CEUeTDCollection
that govern alternatingly, one would not expect that this would lead to decreasing distances between the blocks due to coalition expectations. This openness of coalitions is captured by the measure of government alternation (Casal B´ertoa and Enyedi 2016). The measure ranges from 0 to 100, where 0 means that each time a government changes about half of the composition also changes and 100 means that there is no change in government or that there is complete chance. We could thus expect that lower values of alternation (which on this scale means more alternation) are related to lower polarisation.6
The effect of many of the covariates listed above has been hypothesised to be manifesting primarily or additionally through interactions (e.g. between fragmentation and coalition habits, inequality and turnout and inequality and disproportionality). These interactions will be tested, but they will be included in the model for overall comparison only if they have a notable impact on the quality of the model (model fit).
In order to ensure that all models are comparable to each other, the set of cases is restricted to only those for which data across all the 9 polarisation measures is available. Taking this and the extent of the data that is available for the right hand side variables into account, the coverage of the data set used in the following analysis is 148 elections across 13 countries and is brought out on Figure 5.4. The descriptive statistics of the variables used in the model are brought out in Appendix C.
Data about party system polarisation has both a temporal and a spatial dimension. We have observations for elections within countries over time. Thus, we are dealing with time-series cross-sectional data (TSCS), with a limited number of units repeatedly observed over time. Any analysis of this data should take its specific structure into account, especially the fact that the observations within countries are likely to be related to each other and more similar than across countries. The analysis thus uses the framework suggested by Beck and Katz (1995), Beck and Katz (1996), De Boef and Keele (2008), and Beck and Katz (2011) as the starting point and adopt a model specification that suits the particular characteristics of the data at hand.
Previous research that has looked into various aspects of party system polarisation has been
6 The data that is used in the analysis is obtained from the authors. A version of the measure is used, that is based on a value of alternation for each year. The yearly values are first calculated as follows. If a government changes several times per year, the value for that year is is the average value for all government changes. If there is no government change, the yearly value is 100. In order to reflect the idea that years do not exist in isolation, a weighted average of all the previous values in the party system (weights linearly decreasing to 0 when they reach the beginning of the party system) is used to characterise each year. Such weighted average yearly values for a given election year are used in the current analysis.
CEUeTDCollection
United Kingdom Sweden Spain Portugal Netherlands Italy Ireland Greece Germany France Finland Denmark Austria
1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 5.4: Data Coverage. The figure shows the years for which data for an election (or elections) was available for the listed countries.
heterogeneous both in terms of the data that has been used and partly as a consequence of that, also in terms of the modelling techniques that have been applied. Many of the studies have relied on survey data, especially the comparative study of electoral systems (CSES) (Dalton 2008; Dow 2011;
Curini and Hino 2012; Dejaeghere and Dassonneville 2015). Survey data is limited in the sense that it does not give us observations across many time points and thus it is not possible to model the temporal dimension of the data. Therefore, the studies using this kind of data have relied on simple OLS, perhaps taking into account the clustered and temporal structure of the data by using some form of corrected standard errors.
Using the manifesto dataset, with its extensive temporal coverage, in contrast, opens up possi-bilities for using models that are more in line with the TSCS nature of the data. Analyses that have relied on the manifesto data set, either using the RILE index (Pontusson and Rueda 2008; Matakos, Troumpounis, and Xefteris 2015; Han 2015) or custom dimensions derived from the coding cat-egories (Andrews and Money 2009), have employed models that have taken the spatial (country) heterogeneity as well as the temporal dimension into account. Han (2015) and Andrews and Money (2009) have both used a dynamic model (including a lagged dependent variable in the model) with corrected standard errors. Han (2015) recognizes the fact that country heterogeneity should also be modelled with country fixed effects, but does not use this option as some of the variables that he is interested in are inert or unchanging over time and are thus not possible to include into the fixed effects model. He does use a random effects model for robustness checks, which is theoretically less
CEUeTDCollection
suited for this kind of data (Frees 2004, p. 73; Hsiao 2014, pp. 48-49; see also Beck and Katz 1996;
Clark and Linzer 2015), but models both country differences as well as allows for country invariant predictors. Matakos, Troumpounis, and Xefteris (2015) use both fixed and random effects models, the latter for the same pragmatic reason – to include country invariant variables.
The current analysis will begin with a static fixed effects model, which is somewhere between the most simple, restrictive model – a pooled (no country fixed effects) static model (only contem-poraneous associations at time t) and the most general model, the autoregressive distributed lag (ADL) model, which models both the short and the long term effects of the variables (De Boef and Keele 2008). The analysis will use the static fixed effects model to have a first look at how the two different ways of measuring polarisation perform across the indicators of party difference and position and how this modelling strategy fits the nature of the data. It is checked whether the estimation of country specific intercepts and the inclusion or exclusion of the dynamic component are justified.7