Non-parametric Machine Learning and Explanation in Political Science

Highly predictive machine learning (ML) models such as those based on non-parametric ML effectively predict in complex contexts and are often touted as a way of inductively modelling complex relationships between variables. This often results in these models performing well on prediction tasks when the underlying data-generating process (DGP) contains complex elements. As a result, our use of ML in social science research has soared, with a nearly 300% increase in use from 2017 to 2018-2022.

These models have generally been applied to prediction tasks and are typically considered ill-suited for explanatory tasks in political science for two reasons. First, the most performant prediction models tend to be regarded as `black-boxes.' Second, ML models tend to induce regularisation bias in the pursuit of predictive performance, and thus all estimates derived from these models are subject to regularisation bias.

This PhD focuses on using highly predictive ML models in explanatory contexts in political science. In doing so, I either (1) develop methods to overcome some of the limitations of highly predictive models to facilitate their use in explaining social phenomena or, (2) examine where and how we can best deploy highly predictive models to enhance explanatory research. In the second case, I apply these methods and models to a novel substantive question in political science. The first paper concerns the parameterisation of black-box models with SHAP to generate novel theories. The second paper is an example of leveraging LLMs to facilitate an experimental design, which would not be possible without the capacity to analyse information-dense text data in masse and at speed. My third paper focuses on applying the method and framework I developed in my first paper to the empirical case of Civil War Onset (CWO). Finally, the fourth paper in this PhD examines the potential for using interpretable machine learning methods to enhance explainability in causal inference...