Nason/ice-database

ICE Enforcement Database

A clean, queryable DuckDB database built from ICE enforcement data published by the Deportation Data Project (Berkeley Law / UCLA) via FOIA litigation.

23,568,538 rows across 6 tables covering ICE arrests, detainers, detentions, removals, and custody decisions.

Combines two FOIA releases:

• 2023 release (FY2012-FY2023): arrests, detentions, removals, RCA decisions
• 2025 settlement release (Sep 2023 - Oct 2025): arrests, detainers, detentions

Every row has a data_source column (release_2023 or release_2025) so you can filter by release. Overlapping records are deduplicated, preferring the richer 2025 data.

Built with ice-database.

Quick Start

DuckDB CLI

INSTALL httpfs;
LOAD httpfs;
ATTACH 'https://huggingface.co/datasets/Nason/ice-database/resolve/main/ice.duckdb' AS ice (READ_ONLY);

-- Arrests by month
SELECT DATE_TRUNC('month', apprehension_date) AS month, COUNT(*) AS arrests
FROM ice.arrests
WHERE apprehension_date IS NOT NULL
GROUP BY 1 ORDER BY 1 DESC LIMIT 12;

Python

import duckdb
con = duckdb.connect()
con.sql("INSTALL httpfs; LOAD httpfs;")
con.sql(\"\"\"
    ATTACH 'https://huggingface.co/datasets/Nason/ice-database/resolve/main/ice.duckdb'
    AS ice (READ_ONLY)
\"\"\")
con.sql("SELECT * FROM ice._metadata").show()

DuckDB uses HTTP range requests, so only the pages needed for your query are downloaded.

Tables

Table	Description	Rows	Cols	Sources	Date Range
detentions	Detention stays (book-in to book-out)	9,834,781	41	release_2023, release_2026_08	1995-08-31 to 2026-08-06
removals	Deportation/removal records	3,868,035	43	release_2023, release_2026_08	0001-09-10 to 2026-08-05
rca_decisions	Release/custody assessment decision history	3,543,467	42	release_2023
encounters	ICE encounters with individuals	3,065,104	23	release_2026_08
arrests	ICE administrative arrests	2,516,550	25	release_2023, release_2026_08	2011-10-01 to 2026-08-06
detainers	Detainer requests issued to jails/prisons	740,601	45	release_2026_08	0002-06-14 to 2026-08-05

Key Features

Linked Records

Tables share a unique_id field for tracing individuals across the enforcement pipeline: arrests -> detainers -> detentions -> removals.

Pre-built Views

• v_arrest_to_detention - Arrests joined to detention stays
• v_enforcement_pipeline - Full pipeline: arrest -> detention -> removal
• v_daily_arrests - Daily arrest counts by data source

Multi-release Deduplication

Where both releases cover the same period, records are deduplicated on key fields (unique_id + date + facility) with the richer 2025 release preferred.

Data Source

Deportation Data Project (Berkeley Law / UCLA). Data obtained through FOIA litigation against ICE.

License

Database build code: MIT. Underlying data: public domain (U.S. government records released via FOIA).

GitHub

Full source code, build instructions, and example queries: github.com/ian-nason/ice-database

Changelog

2026-09-14 — Data refresh (Deportation Data Project release of 2026-08-21)

• Rebuilt from the DDP's August 2026 ICE release (activity 2022-10-01 through 2026-08-06), which is cumulative and fully supersedes the March 2026 and December 2025 releases. Total rows 22,030,619 → 23,568,538: arrests 2,516,550 (release_2026_08 share 891,675), detentions 9,834,781, removals 3,868,035, encounters 3,065,104, detainers 740,601, rca_decisions 3,543,467 (unchanged, release_2023 only).
• data_source for the current release is now release_2026_08; the value release_2026 no longer appears.
• New in the source: arrests.city (and county) for the 2022-10+ window (31% populated). ICE did not include apprehension_type in this release.
• Fix: the exact whole-row dedup of encounters (12,419 rows here; 12,097 in July) is now part of the build instead of a one-off repair, and validation asserts zero exact duplicates, view/arrest row parity, no negative durations, and that no superseded release remains.
• Harmonization: Anonymized Unique Identifier, Stay Book Out Date Time and Detention Book In Date Time headers now map onto unique_id, stay_book_out_date and detention_book_in_date (the July build carried a separate, mostly-null stay_book_out_date_time column).
• Builder recognises the DDP's abbreviated file names (ARS/DPR/DET/ENC/REM); scripts/download_releases.sh fetches both releases from their box.com links. DuckDB memory capped at 4 GB / 4 threads; publish_to_hf.py --card-only.
• Caveats unchanged: unique_id never links across releases; detentions is facility-segment level; DDP continues to flag the removals data as less reliable than the July 2025 release.

2026-07-06 — Full refresh + data-quality audit

Rebuilt from the Deportation Data Project's 2026 FOIA release (data through 2026-03-10) and repaired after an independent SQL-verified audit.

Data changes

• New encounters table (2,586,515 rows, release_2026 only).
• Release_2025 fully superseded by release_2026; release_2023 rows inside the 2026 window (2022-10-01 onward) deleted to prevent double counting. Detention supersede keys on stay book-out, removing 17,970 previously double-counted boundary-spanning segments.
• 12,097 exact-duplicate encounters rows removed.
• Totals: 22,030,619 rows across 6 tables (was 17.8M across 5).

Fixes

• Views rewritten: v_arrest_to_detention and v_enforcement_pipeline now produce one row per arrest, matching the first detention stay / removal on or after the arrest within the same release (the old versions joined on unique_id alone, fanning out 3.1x with 20-27% negative durations). New v_detention_stays view collapses facility segments to stays.
• removals.departure_date header rename in the 2026 release handled (previously would load as all-NULL).

Known caveats (see README for the full list)

• unique_id never links across releases (re-anonymized per FOIA release).
• Detainers and encounters exist only from Oct 2022; rca_decisions ends Sep 2023.
• arrests demographics (citizenship, gender, birth year) exist only in release_2026 rows.