Scaling Meaningful Political Dimensions

Item response theory (IRT) models for roll-call voting data provide political scientists with parsimonious descriptions of political actors' relative preferences. However, models using only voting data tend to obscure variation in preferences across different issues due to identification and labeling problems that arise in multidimensional scaling models. Latent Dirichlet Allocation (LDA) models are an increasingly applied approach to using relative term frequencies to estimate the degree to which each text in a corpus discusses a set of issues. However, while models based on relative term frequencies are powerful for discovering which issues are being discussed in which texts, they have proven less useful for discovering variation in political positions within corpuses that cover a range of issues. We combine these two models into a new model for discovering preference variation within issues, using voting data augmented with texts describing each vote. We demonstrate our approach using data from the U.S. Supreme Court.