Three Papers in Applied Political Methodology

• Department of Political and Social Sciences

PhD thesis defence by Haoyu Zhai

How can applied political methodology best help with knowledge accumulation in political science? This thesis presents three related papers in applied methods for political science, each addressing a specific issue in the sub-field: model selection in empirical analysis (Paper I), formal and empirical model integration (Paper II), and model selection and aggregation with a common target (Paper III). The papers share the meta-methodological concern of optimising cumulative learning through enhanced modeling practices across two dimensions of model usage: number (single/multiple) and type (formal/empirical). The main point is that knowledge in political science best accumulates when models are both carefully used within and properly aggregated between empirical studies, so that meta-methodological advances can best benefit meta-learning.

Paper I deals with the issue of model selection in data analysis. It argues that model choice should take into consideration both the key characteristics of the data and the expected gains in model insights. For analysing deliberation data, it recommends the use of network models as such models best respect the complex structure of the data and allow for explicit estimation of endogenous influences driving the process. Paper II looks at the issue of model integration between formal and empirical approaches. It proposes that formal models can be used for guiding not only hypothesis testing but also latent variable measurement. For measuring intraparty unity, it demonstrates how to turn a formal account of faction-based portfolio allocation into an empirical recipe for measuring this latent variable using real-life data. Paper III focuses on a broader issue of model selection and aggregation across multiple studies. It field-tests multiple methods (forecasting and algorithmic) on crowd-sourced statistical models of COVID-19 mortality. For predicting pandemic deaths, it finds that aggregated meta-models generally outperform individual models, and that algorithmic approaches consistently best human experts in achieving such tasks.

Haoyu Zhai is a Research Associate at New York University Abu Dhabi (NYUAD). He has previously been a visiting researcher at the Berlin Social Science Center (WZB) and as a pre-doc at the Nuffield Centre for Experimental Social Sciences at Oxford (CESS). Haoyu's research interests encompass various aspects of political methodology, computational social science, and social data science. His publications have appeared in the British Journal of Political Science (BJPolS) and Research & Politics (RAP). Haoyu holds an MPhil in Politics from Oxford University and a BSc in Politics and Quantitative Research Methods from Bristol University.