Comparing computational and non-computational methods in party position estimation: Finland, 2003-2019

It is often claimed that computational methods for examining textual data give good enough party position estimates at a fraction of the costs of many non-computational methods. However, the conclusive testing of these claims is still far from fully accomplished. We compare the performance of two computational methods, Wordscores and Wordfish, and four non-computational methods in estimating the political positions of parties in two dimensions, a left-right dimension and a progressive-conservative dimension. Our data comprise electoral party manifestos written in Finnish and published in Finland. The non-computational estimates are composed of the Chapel Hill Expert Survey estimates, the Manifesto Project estimates, estimates deriving from survey-based data on voter perceptions of party positions, and estimates derived from electoral candidates' replies to voting advice application questions. Unlike Wordfish, Wordscores generates relatively well-performing estimates for many of the party positions, but despite this does not offer an even match to the non-computational methods.