civilytics/us-cog-finance

US Census of Governments Finance Corpus

American state and local government finance — revenue, spending, and cash and security holdings — harmonized into continuous series across five decades of changing identifiers, item codes, and file layouts.

• Coverage: FY1967–FY2024 (56 published years, 47,126,671 rows)
• Scope: government types 0–3 — state, county, municipality and township
• Format: Hive-partitioned parquet + manifest.json
• Source: US Census Bureau, Annual Survey of State and Local Government Finances and Census of Governments, Individual Unit Files, FY1967–FY2024

FY1968 and FY1969 have no source data: the published years are a set, not an unbroken span. manifest.json names the gaps, and so does GET /api/v1/.

Special districts (type 4) and independent school districts (type 5) are not included. They are ingested in our pipeline but have not yet been through the testing and issue-resolution process the published types have.

Read it

-- DuckDB, no download
INSTALL httpfs; LOAD httpfs;
SELECT * FROM read_parquet('hf://datasets/civilytics/us-cog-finance/data/long/**/*.parquet',
                    hive_partitioning = true) LIMIT 10;

import pandas as pd
df = pd.read_parquet("hf://datasets/civilytics/us-cog-finance/data/long/year=2022")

arrow::open_dataset("hf://datasets/civilytics/us-cog-finance/data/long")

This is the RAW long table. Harmonization, inflation adjustment, per-capita scaling, and provenance are computed on top of it by the uscogdata R package.

Or query it

The same corpus is served by a free, keyless public API — no account, and every response carries its provenance down to the query that produced it: https://cog-api.civilytics.org/api/v1/

Project page and worked examples: https://www.civilytics.org/cog/

Versioning

main is the rolling latest. The data vintage is recorded in manifest.json (built_at, pipeline_commit, data_vintage); for an immutable reference use the commit SHA (resolve/<sha>/). No named version tags are used.

Provenance and limits

Census reported values are preserved untouched; the harmonization layered on top is an interpretation — versioned, documented, and open to disagreement. Totals are reconciled against the Census Bureau's own published aggregates and cross-checked between independent query paths, but a corpus this size is not hand-verified row by row and we do not claim otherwise. If a figure disagrees with a government's audited financial report, we want to hear about it.

Every year other than a census year is a sample, and the sample varies: small governments drop in and out. Compare across years with that in mind.

Cite

Civilytics Consulting, US Census of Governments Finance Corpus (FY1967–FY2024). Sponsored by the Southern Economic Advancement Project; interpretations are our own.