“Partition Decoupling for Roll Call Data”
Data driven measurements of ideology, centered around Poole and Rosenthal's family of NOMINATE models, seek to derive ideological categorizations and relations between individuals based on public roll call votes. While these methods have been very successful, they are constrained, in part, by their strict sets of assumptions. We approach this problem from a new angle, bringing to bear methods from machine learning to construct a data-driven model for ideology from roll call data that is geometric in nature. We adapt the methodology of the "Partition Decoupling Method," an unsupervised learning technique, to produce a multiscale geometric description of a weighted network associated to the roll call votes. The dominant factors in our analysis form a low (one or two) dimensional representation with secondary factors adding higher dimensional features. In this way, our method supports and extends the work of both Poole-Rosenthal and Heckman-Snyder concerning dimensionality of the action space. When used as a predictive model, this geometric view significanly outperforms spatial models such as DW-NOMINATE and the Heckman-Snyder 6-factor model, both in raw accuracy as well as Aggregate Proportional Reduced Error (APRE).
Scott Pauls is an associate professor of mathematics at Dartmouth College, where he has taught since 2001. His work in applied mathematics focuses on the study of complex systems and the construction and analysis of network models. His recent work includes applications to economics, genetics, neuroscience, and political science.