Finding patterns of co-occurrence: Explorative hierarchical cluster analysis with automated item exclusion

Abstract

In the analysis of communication events such as elections, public debates, business communication or campaigns, contingency analyses have become a widely used tool in recent years. By means of analysis of co-occurrence patterns of words, objects or broader semantic elements, the contents of complex communication events may be summarized to few sub-issues, positions, or journalistic frames.
In most cases these analyses use multidimensional scaling techniques to reduce the complexity of large contingency tables. These techniques, however, require decisions by the researcher at critical points. This renders explorative and inductive analyses almost impossible.
In this paper a method for explorative cluster analysis is presented which is specifically designed to reduce the complexity of large consistency tables and to exclude items adding noise to the overall pattern. As an example, the technique is used to detect slight changes in the media framing of unemployment in Great Britain in autumn 2010.