Essays in Political Methodology

Abstract

This collection of essays studies Bayesian statistical models for data analysis in political science.

The first chapter proposes a nonparametric Bayesian approach that uncovers heterogeneous treatment effects even when moderators are unobserved. The method employs a Dirichlet process mixture model to estimate the distribution of treatment effects, and it is applicable to any setting in which regression models are used for causal inference. An application to a study on resource curse also shows that the method finds the subset of observations for which the monotonicity assumption is likely to hold.

The second chapter proposes a new topic model for analyzing plagiarism. Text is modeled as a mixture of words copied from the plagiarized source and words drawn from a new distribution, and the model provides estimated probabilities of plagiarism for each word. An application to the corpus of preferential trade arrangements shows the utility of the model by describing how the likelihood of plagiarism varies across topics.

The third chapter develops an $n$-gram based topic model for analyzing citations and text. It introduces a latent variable indicating whether each token is in the positive or negative mode and assumes dependence between the mode of citation and its previous terms. An application to the data set of the WTO dispute settlement mechanism shows that the conclusion of an existing study using the same data may hold only in some issues.

The fourth chapter (coauthored with Gabriel Lopez-Moctezuma and Devin Incerti) reexamines resource curse using the Bayesian dynamic linear model (DLM). The DLM models both temporal dependence in the data and the evolution of parameters. The results show that there is a negative relationship between oil income and the level of democracy only after the 1970s, which coincides with the Arab oil embargo and the nationalization of the oil industry in major oil-exporting countries.