Scaling at Scale: Ideal Point Estimation with ‘Big-Data’
Jonathan Paul Olmsted, Princeton University
February 27, 2014
The robust literature on the measurement of ideology in political science includes approaches for estimation within multiple inferential frameworks, with and without parametric assumptions on the underlying process, allowing either static or dynamic preferences over time, and from various sources of manifest data. These measurement methods have been applied to a wide variety of political actors and contexts which produce measures of ‘ideology’ in a common low-dimensional underlying space. However, these current ideal point estimation methods are ill-suited for measurement in a ‘big-data’ setting. To remedy this, we derive the necessary results to apply varia- tional Bayesian inference to the ideal point model. This deterministic, approximate solution is shown to produce comparable results to those from standard estimation strategies. However, unlike these other estimation approaches, solving for the (approximate) posterior distribution is rapid and easily scales to ‘big data’. Inferences from the variational Bayesian approach to ideal point estimation are shown to be equivalent to standard approaches on modestly-sized roll call matrices from recent sessions of the US Congress. Then, the ability of variational inference to scale to big data is demonstrated and contrasted with the performance of standard approaches.
Download Dissertation: early draft (about 3MB)