Reproducibility Guide

This project is a modern replication of Carruthers & Ulfarsson (2003), “Urban Sprawl and the Cost of Public Services,” which studied how urban development patterns relate to per capita public expenditure across metropolitan counties in 14 fast-growing US states. The original study used 1982–1992 data; this replication updates the analysis to the 2012–2022 period using Census Bureau data.

Key finding: The density–spending relationship identified in the original paper is specification-sensitive. Whether higher density is associated with lower per capita costs depends on the choice of fixed effects, the treatment of zero-expenditure observations, the inclusion of revenue controls, and the addition of demographic demand variables. The 6-specification ladder reveals which results are robust and which are artifacts of particular modeling choices.

What this repository produces

• Analysis panel: A county-year panel dataset (1,148 observations across 394 counties, 74 columns) combining expenditure, revenue, demographic, employment, land area, and governance data for metropolitan counties in 14 states across 2012, 2017, and 2022.
• 104 regressions: 72 main regressions (6 specifications × 12 expenditure categories) plus 32 robustness checks (4 checks × 4 categories × 2 specs).
• Interactive dashboard: A self-contained HTML dashboard with coefficient comparison charts, specification details, and robustness results.

Prerequisites

• Python 3.12+ — Required for the analysis scripts. The project uses modern Python features (type unions, match statements).
• uv package manager — All Python commands must be run through uv. Never install packages to system Python or use pip directly.
• Census API key — Required for fetching data from the Census Bureau APIs (ACS 5-Year, County Business Patterns).

Get a Census API key

Request a free API key from the Census Bureau at https://api.census.gov/data/key_signup.html. The key will be emailed to you within minutes.

Installation

The pipeline consists of numbered scripts that must be run in order. Each stage reads from the previous stage’s output and writes to a well-defined location.

Stage 1: Fetch raw data (Scripts 01–06)

Scripts 01 through 06 fetch raw data from Census Bureau APIs and bulk download files. All raw data is saved to data/raw/. This stage requires an active internet connection and a valid CENSUS_API_KEY.

• 01_fetch_demographics.py — ACS population, poverty, age composition, property values
• 02_fetch_employment.py — County Business Patterns employment data
• 03_fetch_finance.py — Census of Governments expenditure and revenue files
• 04_fetch_urban_area.py — UA-County relationship files (2010 and 2020 vintages)
• 05_fetch_land_area.py — Census Gazetteer total land area files
• 06_fetch_governance.py — Census of Governments organization files (government counts)

Stage 2: Build panel (Script 10)

Script 10_build_panel.py merges all raw data sources into a single analysis panel at data/final/analysis_panel.parquet. The panel contains approximately 1,148 rows (county-year observations) and 74 columns. CPI deflation is applied before per-capita computation. Government finance amounts (stored in thousands in the raw files) are multiplied by 1,000 during this step.

Stage 3: Run regressions (Script 20)

Script 20_run_regressions.py estimates all 104 regressions and writes results to data/final/regression_results.json. This includes 72 main regressions (6 specifications × 12 expenditure categories) and 32 robustness checks (4 checks × 4 categories × 2 specs).

Stage 4: Build dashboard (Script 30)

Script 30_build_dashboard.py reads the regression results JSON and generates a self-contained interactive HTML dashboard at output/dashboard.html.

Quick Start

The analysis panel (data/final/analysis_panel.parquet) contains 74 columns grouped by type. All dollar-denominated variables are in real 2022 dollars (CPI-deflated). Per capita variables are computed as value / population.

Identifiers

Column	Type	Description
state_fips	str	2-digit state FIPS code (zero-padded)
county_fips	str	3-digit county FIPS code (zero-padded)
year	int	Panel year (2012, 2017, or 2022)
county_id	str	5-digit combined FIPS (state_fips + county_fips)
county_name	str	County name
state_name	str	State name

Expenditure per capita (12 categories)

Column	Description
per_capita_total_direct	Total direct expenditure per capita
per_capita_capital_facilities	Capital facilities expenditure per capita
per_capita_roadways	Roadways expenditure per capita
per_capita_other_transportation	Other transportation expenditure per capita
per_capita_sewerage	Sewerage expenditure per capita
per_capita_trash_collection	Trash collection expenditure per capita
per_capita_housing_community_dev	Housing and community development expenditure per capita
per_capita_police_protection	Police protection expenditure per capita
per_capita_fire_protection	Fire protection expenditure per capita
per_capita_parks_recreation	Parks and recreation expenditure per capita
per_capita_education	Education expenditure per capita
per_capita_libraries	Libraries expenditure per capita

Revenue per capita (2)

Column	Description
per_capita_local_tax_revenue	Local tax revenue per capita
per_capita_intergov_revenue	Intergovernmental revenue per capita

Density & land

Column	Description
density	(population + employment) / urban_land_acres
urban_land_acres	Urbanized land area in acres (time-varying: 2010 or 2020 UA vintage)
total_land_area_sqmi	Total county land area in square miles
pct_urbanized	Fraction of county land area that is urbanized

Demographics

Column	Description
population	Total population (ACS 5-Year)
employment	Total employment (County Business Patterns)
poverty_rate	Share of population below the poverty line
pct_under_18	Share of population under 18 years of age
pct_65_plus	Share of population aged 65 and over

Property

Column	Description
agg_property_value	Aggregate owner-occupied property value (real 2022$)
property_value_per_acre	Aggregate property value / urban_land_acres

Governance

Column	Description
municipal_count	Number of municipal governments in the county
special_district_count	Number of special district governments in the county
per_capita_municipal_govts	Municipal governments per 1,000 residents
per_capita_special_districts	Special districts per 1,000 residents
central_city	1 if county contains a CBSA principal city, 0 otherwise

Fixed effects

Column	Description
state_AZ, state_CA, state_CO, state_FL, state_GA, state_ID, state_NV, state_NM, state_NC, state_OR, state_TN, state_UT, state_WA	13 state dummy variables (TX omitted as reference category)
year_2017	Year dummy for 2017 (2012 omitted as reference)
year_2022	Year dummy for 2022 (2012 omitted as reference)

We estimate six progressively different specifications for each of the 12 expenditure categories (72 regressions total). This “spec ladder” approach reveals how sensitive the density–spending relationship is to modeling choices and helps distinguish robust findings from specification-dependent artifacts.

Summary comparison

Spec	FE	SE	DV	Revenue	Demographics
1	State	HC1	ln(y)	Yes	No
2	County	Cluster	ln(y)	Yes	No
3	County	Cluster	ln(y)	Yes	Yes
4	County	Cluster	ln(1+y)	Yes	Yes
5	County	Cluster	ln(y)	No	Yes
6	State	HC1	ln(y)	Yes	Yes

Four robustness checks are applied to two key specifications (Spec 1: state FE + HC1, and Spec 3: county FE + clustered SEs + demographic controls) across four expenditure categories central to the sprawl–cost debate: Total Direct, Roadways, Sewerage, and Police Protection. Each check produces 8 regressions (4 categories × 2 specs), for a total of 32 robustness regressions.

Running the test suite

The project includes 393 automated tests that collectively verify every stage of the pipeline. All tests can run without network access or a Census API key because they use fixture data stored in data/fixtures/.

What the tests cover

Test Area	Description
Data fetching fixtures	Each fetch script (01–06) has tests that use small sample data in data/fixtures/ to verify parsing logic, column naming, FIPS code handling, and data type conversions without making network calls.
Panel construction	Tests verify that the merge logic in script 10 correctly joins all data sources, applies CPI deflation, computes per-capita variables, creates density measures, and generates fixed-effect dummies.
Regression coefficient matching	Tests verify that the regression engine produces the correct coefficient estimates, standard errors, and p-values by comparing against known results from small fixture datasets.
HTML output validation	Tests verify that the dashboard and methodology page generators produce valid HTML with the expected structure, section IDs, and content.

TDD workflow

This project follows a test-driven development workflow. For every module:

1. Write tests first in tests/test_<module>.py
2. Create test fixtures in data/fixtures/ (small sample data, no network needed)
3. Implement the module in scripts/<module>.py
4. Verify: uv run pytest tests/test_<module>.py -v

Adding a new specification

To add a 7th (or Nth) specification to the ladder:

1. Open scripts/run_regressions.py and locate the SPEC_CONFIGS dictionary.
2. Add a new entry following the existing pattern. Each spec config defines: the fixed-effects type (state or county), the standard error type (HC1 or clustered), the dependent variable transform (ln or log1p), whether revenue variables are included, and which extra control variables to add.
3. Run uv run python scripts/20_run_regressions.py to re-estimate all regressions including the new specification.
4. Run uv run python scripts/30_build_dashboard.py to regenerate the dashboard with the new spec’s results.

Adding new variables

To incorporate a new independent variable into the analysis:

1. Update config.py: Add the new variable’s API codes, item codes, or file references to the appropriate mapping dictionary.
2. Update the fetch script: If the variable comes from a new data source, create a new fetch script. If it comes from an existing source (e.g., ACS), modify the relevant fetch script to include the new API variable.
3. Update scripts/build_panel.py: Add the merge and computation logic for the new variable in the panel construction step.
4. Update scripts/run_regressions.py: Add the new variable to the appropriate specification configs in SPEC_CONFIGS.
5. Write tests for each modified module before implementing changes.

Changing the sample

The 14-state sample is defined in config.py in the STATES dictionary, which maps state abbreviations to FIPS codes. To change the sample:

1. Edit the STATES dictionary in config.py to add or remove states (with their FIPS codes).
2. If adding states, update the state fixed-effect dummy variables accordingly (remember to keep one state omitted as the reference category).
3. Re-run the full pipeline from script 01 through script 30 to fetch data for the new states and rebuild all outputs.