Party positions from different sources – How to combine them with voter data?

Estimation techniques for party positions rely largely on three different types of data sources: political text, i.e. manifestos, expert placements, or voter perceptions as reported in survey data. While the comparative literature on party systems mostly draws on both text-based and expert placements to analyse party configurations, there are at least two broad strands of literature that face the challenge to combine party positions with voter preference data: the literature on policy voting, and on representation. Here, researchers often rely on voters’ perceptions of parties’ policy positions, as reported in surveys. However, this kind of data is scarce, and often only available for the common left-right dimension, but not for more concrete policy scales. Also, it is known that voter preference and perception data on policy issues suffers from biases caused by persuasion and projection effects, as well as from differential item functioning. While expert placements of parties are of very good quality, it is unclear how they fit into voter preference scales. The same holds true for party positions based on manifestos. This data source is additionally criticized for being biased by issue salience. However, manifestos are often the only available database to estimate party positions for a wide range of countries prior to the 1990s, when e.g. the Chapel Hill Expert Survey was first fielded. This paper discusses the combination of party positions from different sources with voter preference scales. We draw on data from the German longitudinal election study and compare the performance of different estimates of party positions that we frequently find in the literature: first, the mean self-placement of a party’s voters, second, the mean expert party placement based on an external survey, and third, party positions based on manifesto data. Additionally, we propose a simple rescaling technique to fit party positions from external sources onto the voter distribution. The benchmark for our comparison is the mean perception of a party’s issue positions based on respondent data from the same survey. Since we are interested in the usage of party positions for spatial vote models, our measure for the goodness of fit is the percentage of correctly predicted vote choices in a conditional logit model of vote choice based on issue distance. Our rescaling technique performs very well with expert survey data, and even outperforms mean perceived party positions from the same survey. Thus, it can be regarded a good substitute if party positions based on perception data are unavailable. This offers many opportunities for comparative researchers interested in policy voting and representation, who are often limited in their research by data availability. Our results raise more general questions about how we should conceptualize and interpret party positions in models of political competition, as well as in representation research. We detect significant differences in the party configuration based on different estimation strategies. Thus, the choice of which party position estimate to use might make a difference. Our research provides an overview of the potential bias that the choice of data source might produce.