Nonparametric and nonlinear panel data and time series analysis
| Associates: |
|Duration:||01.01.2018 - 31.12.2021|
In the recent past, the analysis of network structures has become popular in political science. Specifically, it is of interest to assess how persistent network structures are, to identify key actors driving the political agenda and to study their dynamic interaction. However, when dealing with political actors such as members of parliament or parties, it may not be obvious how to identify relevant relations between them from unstructured data, e.g. text, and how to combine information from heterogeneous sources, e.g. political speeches and manifestos, within a unified network. Our assumption is that hidden relations between political actors can be uncovered by developing a methodology that automatically induces a time series of networks from large quantities of textual content and by analyzing the complex dynamics of such representations with appropriate time series methods. Consequently, the project’s aim is to contribute novel methods along two complementary lines of research: (1) the automatic extraction of networks from political text using natural language processing (NLP) methods and (2) the development of methods for high-dimensional time series that are suitable for analyzing the dynamics of networks formed by political actors. To implement this approach, Simone Ponzetto will join the project as principle investigator and bring his expertise on NLP methods. To describe relations between political actors, we will automatically acquire networks from large collections of political discussions. Given textual content, we can generate networks of increasing complexity and granularity. At the coarsest-level, we investigate turn-based networks, capturing the high-level interaction between the actors, e.g. who addressed whom in a particular meeting. Next, we will look at ways to exploit textual content directly in order to estimate the strength of association encoded within the networks’ weights from a variety of heterogeneous signals and build finer-grained, semantically more informed networks. To this end, we will (i) devise measures that capture topic-specific similarities between actors, employing models for topical analysis, (ii) derive networks’ weights as distances in ideological positions, using recent advances to scale political texts, and (iii) build argumentation-based interaction networks by analyzing actors’ interaction in terms of identified common arguments. The network dynamics will be modeled using the vector autoregressive (VAR) framework. Precisely, we will vectorize the adjacency matrices that describe the network and study the temporal (cross)-dependence of network (non-)links in order to assess network persistency and the dynamic interactions between actors. However, we will face a number of challenges. First, we need to identify appropriate (non-standard) VARtype approaches, such as vector-valued binary time series models for non-weighted networks, and devise corresponding models to handle weighted networks described by non-binary (bounded continuous) outcome variables. Second, as network models are characterized by a large number of parameters, suitable shrinkage estimation methods have to be proposed. Third, we need to explore how typical VAR inference methods such as Granger-causality or impulse response analysis help in understanding the dynamics of networks. For our investigation, we will focus on the transcripts of European Parliament speeches and statements from numerous plenary and committee sittings as the primary data source. European Parliament interactions provide us with a goldmine of data that involve various types of political actors. Additionally, we will continue to work on a complementary setting in close cooperation with project C4, namely the analysis of party manifesto data as a means of studying the evolution of party positions in a political spectrum.