Zoning in American Cities: Are Reforms Making a Difference? An AI-based Analysis

TL;DR

The study investigates whether form-based codes (FBCs) promote sustainable urban form by applying NLP to 2,723 U.S. zoning documents to quantify FBC adoption relative to a reference set of 70 codes using BigBird embeddings and cosine similarity ($FBC\_similarity$). It then links this similarity to urban form and mobility outcomes (setbacks, FAR, minimum plot size, walkability, commute distance, and multi-family housing share) via place-level regressions, finding that higher $FBC\_similarity$ is associated with denser, more pedestrian-friendly configurations, especially in neighborhoods developed after 1950. The analysis also shows that $WRLURI$ regulatory stringency is weakly related to FBC similarity and that regional variation is outweighed by within-region differences, highlighting strong local determinants. Validation with human coders on setbacks and FAR demonstrates reasonable alignment with the NLP-derived classifications, while the authors discuss limitations related to model-based text interpretation and measurement choices and call for further refinement and causal inference approaches.

Abstract

Cities are at the forefront of addressing global sustainability challenges, particularly those exacerbated by climate change. Traditional zoning codes, which often segregate land uses, have been linked to increased vehicular dependence, urban sprawl, and social disconnection, undermining broader social and environmental sustainability objectives. This study investigates the adoption and impact of form-based codes (FBCs), which aim to promote sustainable, compact, and mixed-use urban forms as a solution to these issues. Using Natural Language Processing (NLP) techniques, we analyzed zoning documents from over 2000 U.S. census-designated places to identify linguistic patterns indicative of FBC principles. Our findings reveal widespread adoption of FBCs across the country, with notable variations within regions. FBCs are associated with higher floor-to-area ratios, narrower and more consistent street setbacks, and smaller plots. We also find that places with FBCs have improved walkability, shorter commutes, and a higher share of multi-family housing. Our findings highlight the utility of NLP for evaluating zoning codes and underscore the potential benefits of form-based zoning reforms for enhancing urban sustainability.

Figures (7)

• Figure 1: Differences in thematic content. Panel (a) PCA on document embeddings, showing separate embeddings for the reference set and for FBC codes obtained from our two repositories. Panel (b) FBC similarity distributions for zoning documents (N=2,723) and reference set documents (N=70).
• Figure 2: Word clouds for places with top and bottom FBC. The figure illustrates the prominent themes extracted from zoning documents of top and bottom cities. Panel (a) represents themes prevalent in top FBC cities (top 20% of FBC distribution), panel (b) shows the themes from bottom FBC cities (bottom 80% of FBC distribution), and panel (c) shows the themes for the FBC repository documents. Themes were extracted using the ChatGPT-4 natural language model. A TF-IDF vectorizer was used to identify significant themes, and the difference in TF-IDF scores between top and bottom cities was used to filter the themes displayed. The size of each word reflects its distinctiveness within the group, as determined by TF-IDF, which accounts for both frequency and uniqueness relative to the entire corpus. The colors are included for visual clarity and do not carry analytical significance. Common words were excluded to highlight unique themes in each category. See section \ref{['sec:method']} for a full description of the prompts used to obtain the themes from the documents.
• Figure 3: Relationship between FBC similarity and Wharton Regulatory Index (WRLURI). The figure is divided into four quadrants based on the median values of WRLURI and log FBC similarity. Each point represents a place, with colors indicating the quadrant: high WRLURI-high FBC (green), high WRLURI-low FBC (blue), low WRLURI-low FBC (orange), and low WRLURI-high FBC (red). Dashed lines represent the median values dividing the quadrants. Cities with high populations within each quadrant are labeled. The Pearson correlation coefficient between WRLURI and log FBC similarity is 0.02, indicating a weak relationship. Quadrant counts are as follows: high WRLURI-high FBC (N=248), high WRLURI-low FBC (N=227), low WRLURI-low FBC (N=248), and low WRLURI-high FBC (N=227).
• Figure 4: FBC similarity in the United States. Panel (a) shows FBC similarity for all census places in the United States (N=2,723), highlighting 3 cities with high population in each FBC similarity quantile. Panel (b) plots the distribution of FBC similarity by region.
• Figure 5: OLS estimates of high FBC and morphological measures of urban form. The figure plots point estimates from regressing urban form outcomes on high-FBC (top 20% of the FBC distribution), using five different specifications. The unit of observation is the census place. Panel (a) plots estimates for median street setbacks. Panel (b) plots estimates for street setback deviation. Panel (c) shows the estimates for log floor-to-area ratio (FAR). Panel (d) plots estimates for minimum plot size. Each figure plots the point estimate from these regressions with 95% confidence intervals (point estimate ± 1.96 * SE). Specification I includes state fixed effects, Specification II adds location controls (latitude-longitude), and Specification III further controls for log area (km²) and type of place (borough, city, town, village). Specification IV incorporates zoning vintage dummies (1982-1996, 1996-2008, 2008-2016, 2016-2021). Specification V reports estimates for outcomes computed for neighborhoods developed after 1950 in each place. The number of observations for each outcome is 2,452 for median street setbacks and street setback deviations, 2,450 for log FAR, and 2,452 for log minimum plot size. See Table \ref{['table: urban form discrete']} for point estimates and confidence intervals.