us-shapefiles

Reference geography files for the Deportation Data Project: states, counties, and Census places, each built as a single sf object — attributes and full-resolution geometry together — and published as Parquet with native geometry types (data/<level>-sf.parquet, written by GDAL's Parquet driver with USE_PARQUET_GEO_TYPES=ONLY): read with sf::st_read() (R, GDAL >= 3.12), geopandas, or DuckDB.

Levels

Level	Source	Columns (+ geometry)
states	2025 CB 500k	geoid, stusps, name
counties	2025 CB 500k	geoid, stusps, name
places	2025 CB 500k; pop20 from the block table	geoid, stusps, name, pop20

Conventions: WGS 84 (EPSG:4326); the primary geometry column is geometry, at the source's full resolution (consumers that tile these files simplify per zoom themselves); a level needing a second version of its geometry carries it as an additional geometry column in the same file. pop20 is the place's 2020 Census population on its 2025 boundary (0 for the island territories, whose 2020 blocks carry no population counts).

Naming

County and place names come from the Census NAMELSAD field: counties drop the " County" suffix (any type suffix in the island territories) and places drop any parenthetical (county_name() / place_name() in code/utils.R). The ZIP crosswalk carries the same names, so the published EOIR data joins these files by normalized name — norm_geo_key() in code/utils.R, the normalizer the deportationdata apps apply at runtime. Where two distinct places normalize to the same (state, name) key, the more populous one is kept.

ZIP crosswalk

data/zip-xwalk.parquet maps each 2020 Census ZCTA to a county and place — zcta, stusps, county, county_geoid, place, place_geoid — no geometry. county_geoid/place_geoid join counties-sf.parquet / places-sf.parquet directly, so this file needs no name normalization, except for the rare place_geoid naming a same-named CDP the places build's (state, name) dedup dropped for a more populous twin (see Naming) — join those by name instead. Built by code/zip_xwalk.R from the block table: each ZCTA goes to the county and the place holding the largest share of its 2020 population, or of its land area where it has no 2020 population (including the island territories). Blocks in no place compete per state; a ZCTA whose largest share is in no place has place/place_geoid as NA.

Block table

data/blocks-xwalk.parquet (code/blocks.R) has one row per 2020 Census tabulation block — geoid, pop20, housing20, lat, lon, aland, zcta, county_geoid, place_geoid, cousub_geoid, place_split — with the block's 2020 population and its 2025 county, place and county subdivision from the TIGER/Line FACES files. A block split by a boundary change since 2020 goes whole to the side holding most of its area (place_split marks those). pop20 is NA for the island territories. The places build and the ZIP crosswalk aggregate this table.

County subdivisions, the place/subdivision partition, and GNIS

• data/cousubs-sf.parquet — 2025 CB 500k county subdivisions (towns, townships, CCDs, barrios), every state and territory: geoid, stusps, name, namelsad, lsad, county_geoid, aland.
• data/places-cousubs-sf.parquet — a single non-overlapping partition of the country into cells built from places and county subdivisions (code/places-cousubs.R). Columns: geoid, stusps, name, namelsad, common_name, geo_level (place, cousub, remainder), county_geoid (NA for place cells). common_name is the name to show for a cell: a place's common_name from the xwalk; a municipality-level subdivision (a city anywhere, a town in the six New England states) by its plain name, unless two such in the state share it; any other subdivision or remainder by its namelsad. Town-state cells are subdivisions, with two kinds of place cell: the 12 PA municipalities split by a county line (one cell each), and CDPs that straddle subdivisions (carved out, leaving the subdivisions as remainders). Join data to cells through the xwalk, not by geoid directly. The partition is exact under planar geometry (sf_use_s2(FALSE)), which is how it was built; join points to it with s2 off, as code/gnis-places-cousubs.R does, or a point on a shared edge can land in two cells or none.
• data/places-cousubs-xwalk.parquet — every place and county subdivision geoid with the cell it belongs to (cell_geoid, cell_level, cell_namelsad, cell_common_name) and the rule that put it there (merged across county line for the county pieces of a split PA municipality); cell_geoid is NA only for subdivisions covered by several places. common_name strips the consolidated-government suffixes from place names.
• data/gnis-populated-places-sf.parquet — USGS GNIS populated places as points (feature_id, stusps, name, county_name), from the pinned inputs/DomesticNames_National_Text_<pubdate>.zip; "(historical)" entries and unknown (0, 0) locations are dropped.
• data/gnis-places-cousubs-xwalk.parquet — each GNIS point's cell (cell_geoid, cell_stusps, cell_level, cell_namelsad) and how it was found (match): cell for a point inside a cell; tiger for one off the generalized shoreline but inside the unclipped TIGER/Line place or subdivision; nearest for one TIGER puts in water within 500 m of a cell; NA (nine points) for the rest, which are across the Mexican border or at sea with bad coordinates. cell_stusps can differ from stusps near a state line.

Build everything in dependency order with ./build.sh, or name the scripts: ./build.sh cousubs places-cousubs.

Layout

• code/ — one build script per output (blocks.R, states.R, counties.R, places.R, zip_xwalk.R, ...) and shared helpers (utils.R, where GEO_VINTAGE sets the Census vintage). Run from the repo root: Rscript code/states.R, or ./build.sh. Packages: sf (GDAL with the Parquet driver), arrow, tigris, dplyr, purrr, readr, stringr, glue, mirai, curl, rappdirs.
• inputs/ — the pinned USGS GNIS zip. Census files download at build time and are not stored: cartographic boundaries via tigris, and the TIGER/Line block and FACES zips (~20 GB) into the tigris cache dir or US_SHAPEFILES_CACHE.
• data/ — the outputs (Git LFS).