In the article, we address one of the substantive questions of political science research, namely what the similarities and differences in political parties’ electoral programmes content are. Throughout the analyses we discuss and demonstrate the benefits of Clustering methods in providing methodologically superior solutions of multilevel comparisons dilemmas in large data sets.
In Europe electoral manifestos have been systematically collected for decades, their content has been coded and the data are available (project MARPOR, previously MRP and CMP). Each parliamentary party programme is characterised according to more than a hundred codes that can be merged in seven mutually exclusive and theoretically exhaustive policy domains. In addition, for each document contextual data are recorded: political party, party family, country and election year.
Drawing on the MARPOR dataset, comparisons are commonly carried out between national political parties competing for votes in an election year. On the other hand, countries and party families are frequently compared on the basis of average policy domains’ shares, calculated for different time intervals. In our study we begin with comparisons between countries (aggregates of parties) in the period between 1990 and 2003 through graphically presented profiles (line graphs) and hierarchical agglomerative clustering (dendrogram) to get a big picture and particularly to recognise similarities and differences between Slovenia and other fifty countries. Further, we employ Euclidean distances to graphically present relations between Slovenia and all other countries in a Galaxy, i.e. a graphical format developed for that purpose (the central country is depicted as the Sun and all other countries as different planets allocated around the Sun proportionally to Euclidean distances). In the second step we lower the level of analysis to take differences inside countries into account and cluster individual party manifestos (1300 units) into groups (ideal types) using hierarchical agglomerative clustering methods and K-means method. In a contingency table the obtained groups are split according to manifestos’ origin to estimate type variation inside countries. It turned out that in most countries manifestos belong to the same type (countries as aggregates make sense) but in some countries manifestos (parties) are clearly split between types (countries as aggregates don’t make sense). Additionally, types are to a certain degree associated with party families but independent of election year.
Considering everything, clustering approach has enabled us to simultaneously interpret party manifestos as individual documents as well as country, family or type representatives (members) and estimate the level of similarity between units of interest on each level and in general.