Text-as-Data: Methodological Advances and Applications in Economics

Abstract

This thesis focuses on text in various formats as a primary data source. One part of the thesis focuses on topic modelling as one of the well-established unsupervised techniques widely used to gain initial insights into the latent structures of a text collection. The papers in this section deal with aspects such as the impact of text preprocessing and sampling uncertainty of the results of topic modelling, as well as the identification of an optimal number of topics. Studies from the methodological part of this thesis draw attention to important aspects such as reliability and robustness of topic modelling results and provide practical recommendations for LDA applications. The second part of this thesis deals with text-based indicators in an economic context. For example, one paper examines the relationships between scientific trends in publications in Germany and Poland and real economic indicators. Another paper focuses on the construction of economic policy uncertainty indices for three different countries based on news data. Finally, two papers deal with fiscal policy sentiment and disagreement about fiscal policy in Germany. Here, parliamentary debates in the German Bundestag are the main source for the construction of sentiment indices. Altogether, the results of the application part underline the usefulness of text-based methods for understanding economic dynamics and pave the way for future research in this area.