UC Davis

How do individuals conceptualize and operationalize psychological and political phenomena to inform their opinions, beliefs, and behavior? How do we best interrogate these questions and how do new machine learning methodologies inform and alter this research? In three chapters I offer a first attempt at leveraging these methods to answer substantive questions, introducing a new method, and providing guidelines for researchers on how to use these methods for their own research. In the first chapter, I and my coauthors explore the psychological foundations of policy responses to partisan endorsements/cues. Using the combination of a novel survey experiment, multidimensional scaling, and the causal forest method we evaluate the competing importance of motivated bias and affect transfer. This intersection of design, unsupervised machine learning, and causal machine learning leads us to divergent results, compared to the previous literature. We find that affect, not ideology, is the sole moderator of out-party partisan cueing processes. In my second chapter, we introduce a new machine learning method, with accompanying software, for detecting randomization issues in experiments. Leveraging the predictive power of machine learning in combination with the inferential power of permutations, we provide an approach that vastly outperforms standard methods (e.g., balance tables and F-tests). We illustrate this superior performance across simulated, real, and even fabricated data. Importantly, our method requires no ex-ante model-specification and successfully identifies multi-dimensional covariate imbalance where standard methods do not. While chapters one and two highlight the distinct benefits of adopting and integrating new causal machine learning methods, there exists little-to-no guidance on how and when any given researcher should use these methods. Consequently, in my third chapter I introduce principled guidelines for the application of causal machine learning to experiments. This includes design and analysis considerations, especially in relation to the detection and calculation of heterogeneous treatment effects. The goal of this chapter is to reduce the confusion surrounding these methods and their black-box nature and to illustrate the significant benefits they can provide to researchers when used in a principled, transparent way. Overall, these three chapters contribute to the literature on both political behavior and experimental research by introducing, integrating, and applying causal machine learning to experiments in political science.