The Costs of Housing Regulation: Evidence From Generative Regulatory Measurement

Abstract

We present a novel method called "generative regulatory measurement" that uses Large Language Models (LLMs) to interpret administrative documents. We demonstrate its effectiveness in analyzing municipal zoning codes, achieving 96% accuracy in binary classification tasks and a 0.87 correlation for continuous questions. Applying this approach to a comprehensive sample of U.S. zoning regulations, we establish four facts about American zoning: (1) Housing regulations are multidimensional and can be clustered into two main principal components. (2) The first of which corresponds to value capture, indicating how municipalities extract economic benefits in areas of high housing demand. (3) The second principal component associates with exclusionary zoning, resulting in higher housing costs and socioeconomic exclusion. (4) Zoning follows a monocentric pattern with regional variations, with suburban regulations particularly strict in the Northeast. We develop a model of non-cooperative municipal government regulatory choice consistent with these facts.

Keywords: housing regulation, zoning codes, large language models, natural language processing, artificial intelligence, municipal ordinances, retrieval augmented generation

JEL Classification: R52, R58, K11, O38, R31, C81