At the final meeting of the semester, the Computational Political Science for Democracy (CPS4D) Working Group presented ongoing research on the use of artificial intelligence and machine learning methods to model democratic development and decline. The presentation featured collaborative work by Michael Coppedge, Dmitry Zaytsev, Valya Kuskova, and collaborators.
The project seeks to improve upon traditional structural equation and path models commonly used in democratization research by employing AI-based methods capable of identifying nonlinear relationships, feedback loops, and complex causal interactions across large sets of political and socioeconomic variables. The broader goal is to develop computational models that can better explain both democratic stability and regime change over time.
The presentation focused on a neural vector autoregression framework trained on cross-national panel data drawn primarily from the Varieties of Democracy (V-Dem) project and related international datasets. The model uses electoral democracy (polyarchy) as the principal outcome variable and examines how institutional, social, and economic factors interact dynamically over time.
A central finding of the analysis is the importance of a “protective belt” surrounding democracy. Variables such as civil society participation, state capacity, and rule of law, and institutionalized political parties appear to reinforce one another and contribute to democratic stability through mutually reinforcing feedback effects. The AI-based model not only confirmed several relationships identified in earlier path analyses, but also uncovered additional feedback dynamics and nonlinear relationships that were not captured in the original framework.
The researchers also highlighted the distinction between variables that are causally influential and variables that are most necessary for accurate forecasting. Some variables with relatively modest causal effects nonetheless proved highly important for predictive performance, particularly when nonlinear or threshold relationships were involved. This distinction, the presenters argued, illustrates how forecasting-oriented approaches can complement more traditional explanatory models in comparative politics.
Another important contribution of the project is methodological. The AI framework automatically differentiates between variables that function primarily as long-term structural causes, such as geography or historical demographic patterns, and variables that are more responsive to short-term political dynamics. This allows the model to generate a layered representation of democratic development, distinguishing between distal, intermediate, and proximate influences.
The session concluded with a discussion of future directions for the project, including recent conference papers and ongoing publications. Participants discussed both the substantive implications of the findings and the broader potential for integrating computational methods into the study of democracy and political development.
