Research

Rhetorical form is politically consequential but difficult to measure at scale. I introduce rhetorical structure embeddings (RSEs), vector representations of rhetorical form derived from adjacent-layer transformations in BERT. Combined with unsupervised dimensionality reduction and clustering, the RSE-based discovery pipeline identifies recurring patterns in how political text is expressed and organized. Controlled perturbation tests show that RSEs change more as edits alter more of a passage’s rhetorical form. In both a general register-recovery task and a topic-restricted variant, the RSE-based pipeline substantially outperforms conventional readability, complexity, and grammatical benchmarks. I then apply the method to U.S. Democratic and Republican national party platforms from 1980 to 2024, finding six recurring rhetorical forms and differences in their relative use across parties. Democrats more often organize paragraphs around specific problems and proposed remedies, while Republicans more often use compact statements centered on broader goals or action-first policy priorities. The method provides a scalable way to discover and analyze rhetorical form without predefined categories or reducing it to a single dimension such as complexity.

Political scientists often rely on proxy networks to study relationships such as patronage, trust, and coalitional alignment that are not directly observed. Yet proxies based on co-working, shared hometowns, schools, or clubs may capture only part of these relationships and be unevenly observed across dyads. We propose Multi-indicator Latent Tie Inference (MuLTI), which combines relational traces and dyadic information to estimate distinct latent ties. We establish nonparametric identification using conditionally independent, tie-specific anchors and estimate the model by maximum likelihood with stratified stochastic optimization. Simulations show that MuLTI outperforms proxy-based benchmarks, especially when indicators are noisy and multiple indicators are available. Applied to 4,424 Chinese Communist revolutionary elites and more than 9.7 million dyads, MuLTI reveals a positive association between wartime adversarial exposure and later Cultural Revolution mortality. MuLTI provides a scalable framework for estimating distinct latent relationships from imperfect sources of information that conventional proxy coding cannot capture.

In post-conflict states, victorious rebel organizations often assume the position of power. How do they distribute authority post war? The existing literature has shown the importance of ideology, battlefield performance, and military rank in explaining which rebels become more successful politicians and administrators in the post-conflict state. We suggest that the social capital accumulated during the war is an important predictor of post-war political career. Using original data and a novel method to elicit a comprehensive network structure from individual biographies, we provide evidence to support this claim from the early period of Communist China. Exploiting plausibly exogenous variation in the social networks of the Communist rebels due to battlefield deaths, we show that rebels who lost social capital had less successful careers in the elite hierarchy of Communist China than those who gained it. The findings show that revolutionary victory does not simply reward wartime service; it converts surviving relational capital into state power while leaving those embedded in shattered wartime communities at a disadvantage.