Ulric B. and Evelyn L. Bray Social Sciences Seminar

Abstract: We introduce Gaussian process regression (GPR) as a powerful and flexible modeling framework for political science research. GPR offers a compelling middle ground between the strict assumptions of traditional linear models and the highly flexible but often opaque nature of many machine learning methods. This paper provides an accessible overview of GPR, explaining its core concepts, the construction of informative kernels, and inference techniques for key quantities of interest. We demonstrate GPR's versatility through four applications, showcasing how it can be leveraged to build models that are both flexible and structured, focusing particularly on scenarios involving time series. All models are implemented using GPyTorch, which allows for modular construction of customized models, exact estimation in small-n settings, and efficient approximate solutions for large-n applications.

Coauthors are Annamaria Prati, Yehu Chen, and Roman Garnett