Download README.md from Nason/sahie-database: direct link, hf CLI and curl.
- Browser
- Download file 8.26 kB
-
https://huggingface.co/datasets/Nason/sahie-database/resolve/main/README.md
- Command line
-
hf download hf://datasets/Nason/sahie-database/README.md
-
curl -L -o README.md https://huggingface.co/datasets/Nason/sahie-database/resolve/main/README.md
pretty_name: Census Small Area Health Insurance Estimates (SAHIE), 2024
license: other
license_name: us-public-domain-census-bureau-work
license_link: https://www2.census.gov/foia/ds_policies/ds027.pdf
tags:
- health-insurance
- census
- sahie
- duckdb
- datapond
size_categories:
- 100K<n<1M
County and State Health Insurance Estimates
Census estimates of health-insurance coverage for U.S. states, DC and counties in 2024. The August 2026 Small Area Health Insurance Estimates (SAHIE) release provides counts, percentages and 90% margins of error for demographic and income groups. Start with county_uninsured for one row per county for people under 65, all races, both sexes and all incomes. These are modeled aggregate estimates; categories overlap and are not individual records.
Start querying
Inspect New York county uninsured estimates and their 90% margins of error. Estimates describe the Census-defined population used for poverty estimates, not the entire resident population; missing values are not zeros.
fips is the geographic identifier, county_name is its label, pctui is percent uninsured, and pctui_moe is the 90% margin of error in percentage points.
Expected first counties are Albany, Allegany and Bronx, with uninsured percentages
4.8, 5.4 and 8.1 and 90% margins of error 0.5, 0.7 and 0.6 percentage points.
Python returns DECIMAL percentages as Decimal; R returns numeric values. FIPS
identifiers remain strings. These are source estimates, not individual records.
Choose R or Python. Install datapond once: use install.packages("datapond") in R, or run python -m pip install datapond in a terminal.
R
library(datapond)
con <- dp_connect("sahie")
tryCatch({
DBI::dbExecute(con, "SET threads=1")
DBI::dbExecute(con, "SET memory_limit='128MiB'")
DBI::dbExecute(con, "SET max_temp_directory_size='0B'")
print(DBI::dbGetQuery(con, "SELECT fips,county_name,pctui,pctui_moe FROM county_uninsured WHERE statefips='36' ORDER BY fips LIMIT 3"))
}, finally = dp_disconnect(con))
Python
import datapond
con = datapond.connect('sahie')
try:
con.execute("SET threads=1")
con.execute("SET memory_limit='128MiB'")
con.execute("SET max_temp_directory_size='0B'")
print(con.execute("""SELECT fips,county_name,pctui,pctui_moe FROM county_uninsured WHERE statefips='36' ORDER BY fips LIMIT 3""").fetchall())
finally:
con.close()
Queries run on your computer and read the byte ranges they need. A small result can still require a large scan. These examples limit DuckDB to one thread and a 128 MiB buffer pool with disk spill disabled; that is not a limit on the entire R or Python process. Larger analyses may need more memory or a local download with enough disk space.
Tables, fields and joins
See the data dictionary for column definitions.
| Table | Description | Rows | Columns |
|---|---|---|---|
estimates |
Demographic/income estimates, overlapping groups | 402,786 | 27 |
geographies |
State and county FIPS/names | 3,195 | 6 |
category_codes |
Official category definitions | 26 | 3 |
county_uninsured selects under-65, all-race, both-sex, all-income estimates for
3,144 counties. Source missing values remain SQL NULL.
Join category fields (agecat, racecat, sexcat, iprcat) to category_codes using the category name and code. fips is two characters for state rows and five for county rows; filter the geography level before joining, and keep leading zeros.
Interpretation
The groups overlap, so table rows must not be added as population counts. Race detail is state-only. Match geographic vintages before FIPS joins. Kalawao County is unavailable. Zero margins of error reflect source rounding. These aggregate modeled data do not identify causal effects. See the field guide, data dictionary and the official layout.
The population denominator is the Census poverty universe. pctui uses the selected income group; pctelig uses the same demographic group at all incomes. Published 90% uncertainty reflects model and sampling error, subject to model assumptions, and excludes nonsampling error. Missing estimates remain NULL. The field guide defines all measures, overlapping categories, zero-denominator missingness, Kalawao omissions, rounding and geographic-vintage caveats.
Other ways to connect
Direct DuckDB SQL
-- In DuckDB with HTTPFS available:
ATTACH 'https://huggingface.co/datasets/Nason/sahie-database/resolve/main/sahie.duckdb' AS sahie (READ_ONLY);
USE sahie;
SELECT fips, county_name, pctui, pctui_moe
FROM county_uninsured WHERE statefips='36' ORDER BY fips LIMIT 3;
Direct hosted access with Python or R
These direct DuckDB examples access the same hosted snapshot as the live catalog.
Install Python duckdb or R duckdb and DBI; INSTALL httpfs below installs
DuckDB's HTTP extension if it is not already present.
import duckdb
con = duckdb.connect(config={"memory_limit": "128MiB", "threads": "1",
"max_temp_directory_size": "0B"})
try:
con.execute("INSTALL httpfs; LOAD httpfs")
con.execute("ATTACH 'https://huggingface.co/datasets/Nason/sahie-database/resolve/main/sahie.duckdb' AS sahie (READ_ONLY)")
print(con.execute("""
SELECT fips, county_name, pctui, pctui_moe
FROM sahie.county_uninsured
WHERE statefips = '36' ORDER BY fips LIMIT 3
""").fetchall())
finally:
con.close()
library(DBI)
con <- dbConnect(duckdb::duckdb())
tryCatch({
dbExecute(con, "SET memory_limit='128MiB'")
dbExecute(con, "SET threads=1")
dbExecute(con, "SET max_temp_directory_size='0B'")
dbExecute(con, "INSTALL httpfs; LOAD httpfs")
dbExecute(con, "ATTACH 'https://huggingface.co/datasets/Nason/sahie-database/resolve/main/sahie.duckdb' AS sahie (READ_ONLY)")
print(dbGetQuery(con, "
SELECT fips, county_name, pctui, pctui_moe
FROM sahie.county_uninsured
WHERE statefips='36' ORDER BY fips LIMIT 3"))
}, finally = dbDisconnect(con, shutdown=TRUE))
The sahie dataset is available in the live Datapond catalog.
Queries run on your machine. The 14.17 MB file size is storage, not required RAM;
DuckDB's 128 MiB limit covers its buffer pool, not the entire Python/R process.
Selected columns reduce transfer, but a final LIMIT does not guarantee a small
scan. These examples use one thread and disable disk spill.
Sources and maintenance
Sources and license
Source: U.S. Census Bureau SAHIE. License: U.S. public domain (Census Bureau work).
Source: Census SAHIE, 2024, retrieved November 2, 2026. U.S. public domain as a Census Bureau work under the agency policy; no worldwide CC0 claim.
Artifact and client checks
The 14,168,064-byte database includes standard datapond metadata.
Source ZIP SHA-256: 5ec0a0cc5cb2e4e17cf046d9f68508decfdb4d9bbb982781d94bbf56ce50347a.
Tested engine versions are Python DuckDB 1.5.6 and R DuckDB 1.5.5.
Python and R clients can query this file directly with DuckDB/DBI. The Python
datapond 0.1.5 and CRAN R datapond 0.1.0 clients passed HTTP range tests against
this exact artifact using a temporary registry (DuckDB 1.5.6 and 1.5.5 respectively).
The sahie identifier is now available in the live Datapond catalog; the first-use
examples above use that default catalog. See build-manifest.json for identity and validation, and
the code repository for client examples.