Measuring political concepts over time with constrained large language models

This dissertation proposes the Measurement, Evaluation and Validation (MEV) framework for constructing historically grounded political variables from text and using them to study long-run social and political change. The central problem addressed is how to generate meaningful political time series from language when discourse, context, and interpretation evolve over time without conflating semantic drift with substantive change. Rather than treating large language models (LLMs) as autonomous predictors, the dissertation embeds them within constrained measurement pipelines that fix interpretive logic ex ante, evaluate procedural reliability, and validate resulting outputs as reusable social-scientific instruments. Two empirical applications demonstrate the framework under distinct inferential conditions. The first analyzes semi-structured open-ended survey responses that permit direct comparison with conventional closed-ended measures. In this setting, the LLM operates within a referent-grounded pipeline that anchors interpretation to historically situated political objects, producing a MEV-variable—Partisan Affect—whose dynamism is reconstructive: historically specific meanings remain accessible behind numerically invariant scores, allowing those values to be reinterpreted without redefining the variable itself. The second application examines weakly structured problem-centered survey text for which no directly comparable baseline exists. Here, the dissertation constructs a composite MEV-variable—emphNational Economic Ideology—through frame-grounded interpretation of economic reasoning across policy domains. In this case, dynamism is configurational: interpretive frames are stabilized rather than referents, enabling ideological change to be analyzed as shifts in how stable components are combined and distributed without requiring semantic reconstruction of the underlying economic concepts. Together, these applications demonstrate that MEV-variables can be both temporally interpretable and analytically reusable, with the form of dynamism determined by the inferential structure of the measurement problem rather than by model flexibility. The contribution of the dissertation lies not in any single empirical finding, but in establishing a disciplined measurement framework that makes longitudinal inference from text conceptually transparent and methodologically defensible.