Will Lowe, Princeton University:
Scaling things we can count
Research in political text analysis has offered many ways to take counts of words, policy claims, or topics, and project them into substantively meaningful spatial representations. In this talk I will argue that existing models for scaling count data are all in fact all versions of, approximations to, or special cases of a single count data unfolding model that produces low dimensional representations of patterns of relative emphasis. By drawing out its similarity to well-known vote scaling models, this unification makes it easier for users of such methods to know when and how to apply them, and clarifies what they are assuming when they do so. By explicitly connecting topic scaling e.g. using hand coded event data to word scaling the unified model also vindicates some, though not all, of the key intuitions of the saliency theory of party positioning. And by virtue of applying to more kinds of count data than text or topics, I show the same model can recover human-coded conflict and cooperation judgements directly from streams of international event counts extracted by computer from newswire. Along the way I introduce and motivate biplots as an important but as yet underused graphical tool for understanding and communicating scaling models.
Will Lowe is a Senior Research Specialist and Lecturer in Public and International Affairs at the Department of Politics, Princeton University. He is a political methodologist specializing in statistical text analysis with applications to international relations, legislative politics and related fields. For more info, visit his website: http://conjugateprior.org/