Democratic Peace

This paper reviews and critiques prior embedding analyses of the democratic peace, proposes an improved methodology that leverages foundational embedding models, and reports replication results. The author evaluates Rathbun, Parker, and Pomeroy’s microstructural work—including surveys, computational studies, and meta-analysis—offers a systematic critique of embedding analysis practices, and introduces a new approach intended to address those issues while taking advantage of foundational embedding models. The replication presented in the abstract accords with the earlier finding that the democratic peace effect is racialized, and the paper frames its contribution as a guide for conducting embedding analysis in political science.

Survey experiments often describe attributes in a hypothetical scenario to learn about those attributes' real-world effects. Those inferences rest on a frequently overlooked assumption: experimental conditions must be information equivalent (IE) with respect to background features of the scenario. IE fails when subjects, upon receiving information about one attribute, update beliefs about other attributes. For example, labeling a country 'a democracy' can change respondents' beliefs about its geographic location. When IE is violated, the measured effect of the manipulation need not equal the target quantity—the causal effect of beliefs about the focal attribute.