Logic of the Analyses

that surrounds the concept”, which would allow to “generate theoretical predictions which, in turn, lead directly to empirical tests involving measures of the concept” (Carmines and Zeller 1979, p. 23).

One of the weak points of this strategy can be that that there are no well developed theories in social sciences (Strauss and Smith 2009, p. 9). But it has also been noted that it is not necessary to have formal and fully developed theories for construct validation, just enough theoretical prior knowledge about a phenomenon to outline theoretically derived hypotheses, i.e. relationships (Carmines and Zeller 1979, p. 24), that one could expect.

If we look at how various measures of party position have been applied in practical research, then even if we do not have strong theories in every case about how the political profiles of parties should be related to other phenomena, there is a broad range of analyses that have tested possible associations and that could be used for comparing the measures to each other. This is certainly true for coalition formation, where there is a strong consensus that political differences should be related to the likelihood of a coalition forming. In other areas like party system polarisation and party political change, there is less consensus over possible associations, but enough of prior research that would allow for comparison. If one measure captures political differences better than another, then, all else being equal, this should be seen in models when we look at model fit or the strength of associations. Thus, through convergent validity it is possible to compare how well various measures capture the political differences between parties. Taking all of the above into account, this is the strategy that is used in three of the four chapters that follow.

right dimension, we could ask how different assessments of distance are in line with people’s actual or intended behaviour. People can place themselves on a left-right dimension and thus we can know how far they are from all of the parties. Or we can ask them how far they are from each party separately, without specifying any dimension. These different assessments of distance can be compared with how likely a person says they are to vote for a party. As far as political considerations play a role in voting, the better measure should be more in line with voting intention.

The last three chapters zoom in on the manifesto data and the index of similarity. Chapter 5 focusses on party system polarisation, Chapter 6 on coalition formation and Chapter 7 on how parties change their political profiles. All of these are phenomena, for which the measurement of difference between parties is essential and their positions on a left-right dimension are of secondary importance. Furthermore, most of the analyses in these three domains have employed thus far the RILE left-right index. Although one could say that research into these three phenomena is at various degrees of maturity – coalition formation has perhaps been most systematically studied and there is most consensus about the explanation of the phenomenon, while polarisation and party change show less certainty over what should be related to the variable of interest and how – we have rather clear theoretical and reasonably clear empirical expectations across the three topics about what a basic model should look like.

Therefore, all of these chapters will follow the same structure. First, the state of the current research is elaborated and core variables, which are related to the variable of interest, are outlined.

Because political differences are among theexplanans in the study of coalition formation, less atten-tion is paid to other possible predictors of coaliatten-tion formaatten-tion. For polarisaatten-tion and change, political differences are the explanandum and therefore one should be careful to outline a more complete model, that would not exclude a crucial explanatory variable, but would yet maintain simplicity that is commensurate to the amount and nature of data that is available for these analyses. For each of the topics, all models are fit to exactly the same set of cases, so that the only thing which differs between the models is the measure of party politics. Any difference in model fit or the associations that are indicated are thus attributable to the latter. All else being equal, the more valid measure should provide the better fitting model.

CEUeTDCollection

Chapter 4

Direct Pairwise Comparisons as a

Means to Understand Political Space

Note: This chapter has been prepared as a separate article, which indicates Andr´e Krouwel as the second author. This is to recognize his role as the academic director of Kieskompas (www.kieskompas.

nl ) for providing the individual level data that is used in this chapter.

He that walketh with wise men shall be wise: but a companion of fools shall be destroyed.

– Proverbs 13:20,The Bible (King James version)

Tell me what company thou keepest, and I’ll tell thee what thou art.

– Miguel de Cervantes / Sancho Panza,Don Quixote

The left-right dimension is the default way of thinking about political or ideological spaces¹ in political science. It is the dimension most common across countries, the prevailing idea one finds across much of the literature on political spaces both in Western (see e.g. Huber and Powell 1994, p. 294; van der Brug 2001, p. 117; Budge and McDonald 2006, p. 253; Benoit and Laver 2006, p. 111;

Dalton 2008, p. 910) as well as non-Western countries (e.g. Singer 2016, p. 180). Even though there is research that argues Western political spaces to be two- (Kriesi et al. 2006; Bornschier 2010b) or (Warwick 2002) three-dimensional, or of varying dimensionality across countries (Benoit and Laver 2006), the left-right dimension is still the one which is most used in research that requires us to assess

1 These two terms are employed interchangeably to refer to the space that is used to represent the political differences between parties in a party system.

CEUeTDCollection

the differences between parties, like in studies of coalition formation (e.g. Martin and Stevenson 2001) or party system polarisation (e.g. Dalton 2008).

If a survey is to include an instrument about party ideology, it is most often about positions on the left-right scale, not because this is the only dimension that is of importance, but most likely because this is the only dimension that political scientists can agree on (Warwick 2002, p. 116). However, using survey tools that just see left and right, we lack an understanding of how the complete political space is structured in the minds of people. This chapter shows how this gap can be filled using pairwise comparisons between parties and multidimensional scaling (MDS). As a survey instrument, it is more demanding than left-right placement, because it requires the evaluation of all party pairs by the respondent, but the results that it can give us are uncontaminated – they do not depend on our a priori assumptions about what issues to ask people or what dimensions to present them with.

The method of comparative judgement together with multidimensional scaling (MDS) is the method of choice for the study of the structure of perceptual spaces (G¨ardenfors 2000). It takes all the individual differences between object pairs and strives to represent these in lower dimensions.

An often cited example to illustrate this is through distances between cities (e.g. Kruskal and Wish 1978). Suppose we had data on how far in a straight line all European capitals were from each other.

Each city is characterised by as many values as there are other capitals. If we use MDS to represent these values in a 2-dimensional space and plot it, we will get essentially a map of Europe, where each city is characterised by two values – longitude and latitude or its location on the x and y axes of the map. A complex data structure comfortably reduces to a simpler form, as the initial pairwise distances originated from a two dimensional space and the pairwise distances that we started with are exactly the same as the distances represented in the two dimensional space.

We can do exactly the same thing with political parties. Survey respondents can be asked to assess the perceived political or ideological differences between parties one against every other. Such pairwise distances can then be analysed with MDS, which enables to evaluate and uncover the underlying perceptual space for the respondents’ pairwise assessments. At the start of the analysis we would not know and we would not need to know how this space looks like. There would be no necessity to define a benchmark for evaluation – in this case an ideological dimension – beforehand.

We can create lower dimensional representations of the pairwise assessments and then ask how well they account for the original data. This method is particularly well suited for the study of political spaces for which it would be problematic to have pre-defined measurement scales or benchmarks without distorting the space we study. For previous applications of this method in political science,

CEUeTDCollection

see section 3.4.1.

The aim of the current chapter is thus to emphasise an overlooked, but potentially very insightful use for the idea that parties should first and foremost, if possible, be compared to each other in pairs. This is a method, which not only gives us a direct estimate of the difference between any two parties, but also provides a useful way to analyse the underlying structure of people’s thinking about political parties. The following demonstrates its feasibility to provide inductive representations of political spaces. The chapter continues with a brief account of MDS and pairwise comparisons, introduces the dataset that has been gathered for Germany, the Netherlands and Sweden and ends with the presentation of the results of several approaches for using MDS with this data. Even though the focus of this chapter is on individual level analysis as here the gap in the current research into party politics is the largest, the results are also compared to analysing the party difference structure in these three countries at approximately the same time on the basis of manifestos. Although there are differences between those two sources, there is also a notable amount of agreement between them. In the end of the chapter, the results are interpreted in the context of some of the recent studies about the dimensionality of political spaces in European countries in general and in these three party systems in particular.