Advancing Political Science With Machine Learning: A Gaussian Process Approach

The proliferation of data in recent decades including including survey, image and text data, has significantly transformed the landscape of political science research. Machine learning methods have played an instrumental role in analyzing these datasets, yet their application poses challenges in areas where political concepts are not directly measurable or the primary focus of inference is causality. In addition, the essence of machine learning algorithms being trained for prediction performance in a black-boxed manner, makes their outputs hardly interpretable and even unappreciated. This dissertation addresses these issues by proposing a novel methodological framework that employs Gaussian Process (GP) models, a non-parametric Bayesian approach that combines the flexibility to model complex, non-linear relationships and interpretability necessary for causal inference. Through a series of advancement in latent variable measurement, causal inference, prediction and experimental design models, this dissertation demonstrates the effectiveness of the GP framework in addressing core quantitative challenges in political science, thereby advancing the methodological toolkit available to researchers in the field.