Statistical Analysis of High-Level Features from State of the Union Addresses

A computational political science approach is taken to analyze the State of the Union Addresses (SUA) from 1790 to 2015. While low-level features, e.g. linguistic characteristics, are commonly used for lexical analysis, the authors herein illustrate the utility of high-level features, e.g. Flesch-Kincaid readability, for knowledge discovery and discrimination between types of speeches. A process is developed and employed to exploit high-level features which employs 1) statistical clustering (k-means) and a literature review to define types of speeches (e.g. written or oral), 2) classification methods via logistic regression to examine the validity of the defined classes, and 3) classifier-based feature selection to determine salient features. Recent interest in the SUA has posited that changes in readability in the SUA are due to declining audience capabilities; however, the authors' results show that changes in readability are a reflection of changes in the SUA delivery medium.