civilytics/crdc-school-arrest-rates

CRDC School Arrest Rates — Bayesian Estimates

Model-based estimates of school-based arrest rates for U.S. school districts (LEAs) and states, by race and sex, derived from the U.S. Department of Education Civil Rights Data Collection (CRDC). Estimates come from Bayesian hierarchical binomial models (brms / Stan) that partially pool sparse counts, producing stabilized rates with full posterior credible intervals.

Data release: civilytics-crdc-arrests-2026.1, a preliminary update. The white paper's release is civilytics-crdc-arrests-2025.1.

Each release is a tag on this dataset, and its files never change once tagged. civilytics-crdc-arrests-2025.1 stays available under its own tag. main holds the latest release, so pin a tag when results need to be reproducible.

Release 2026.1

2026.1 is a preliminary update to 2025.1, the release the white paper describes. 2025.1 was vetted for the paper; 2026.1 has not yet been vetted the same way. Pin civilytics-crdc-arrests-2025.1 for the paper's numbers, cite the white paper (below), and name the release you used.

2026.1 adds the 2023-24 CRDC, and it also changes the 2015-16 to 2021-22 numbers:

• A school that did not report arrests is now missing, not zero. 2025.1 read every CRDC reserve code as zero arrests, so a school under an action plan (-5) or force certification (-6) counted as a school with no arrests. Every code except -9 (not applicable) is now missing, apart from a 2023-24 -3 (a skip-logic failure), which counts as zero where the school reports no arrest instances for that group. A missing school is left out of both the arrest count and the enrollment it is divided by.
• New York City is in. CRDC and the NCES Common Core of Data file the city's schools under different IDs, so 2025.1 lost them in every wave. 2026.1 matches them through the CRDC Appendix K crosswalk. The city is modeled as its 33 geographic districts in 2021-22 and 2023-24, with a citywide total alongside.
• The recent-year models (m1, m2) now fit 2023-24, and the pooled models (m3–m5) pool all four waves.

Two limits of this release, documented in full in the data dictionary:

• Five district intercepts in unified_m3 have R-hat between 1.010 and 1.012, just above the usual 1.01 target. The other nine models reach a largest R-hat of 1.010 to 1.022.
• The pooled models give each district one intercept across all four waves, so a misreported year moves that district's estimate in every wave. DeKalb CUSD 428 (IL) reported 1,464 arrests in 2021-22, equal to its referrals, and 2 in 2023-24. For 2023-24, unified_m3 estimates 311 arrests (95% interval 281 to 345).

What's in this dataset

Path	Description
summary.duckdb	Compact DuckDB (325 MB). arrest_summary (LEA grain, 2,912,972 rows), state_summary (draw-wise population aggregate, 11,424 rows), district_dim (names, geography, and New York City's parent_leaid / is_aggregate), models (the release's model registry) and meta. Powers the live API.
parquet/	The raw posterior draws (500 per group) as Hive-partitioned Parquet, partitioned by model_id / YEAR / LEA_STATE and sorted within shard by (LEAID, RACE, SEX). 1,830 shards, 1.7 GB.
stages/	Pipeline intermediates that rebuild the reports without refitting: the model inputs, per-wave CRDC files, diagnostics and the five unified fits (6.5 GB of the 7 GB). See data stages.

Coverage / code lists

• RACE ∈ {AM, BL, HI, WH} · SEX ∈ {F, M} (8 demographic cells, no TOTAL)
• YEAR ∈ {15-16, 17-18, 21-22, 23-24}. The recent-year models (m1, m2) cover 23-24 only.
• Models: 10 specifications (unified_m1–unified_m5, stratified_m1–stratified_m5); default = unified_m2 / stratified_m2 (most-recent-year + referral-rate covariate).
• Estimates include count + rate point estimates and HPD intervals at 50 / 80 / 95%.
• State summaries are a draw-wise population aggregate (sum across LEAs within each posterior draw), distinct from the model's (1|LEA_STATE) random effect.
• New York City appears twice in 2021-22 and 2023-24: as its 33 geographic districts and as one citywide row, LEAID = 3620580. district_dim marks them: each district has parent_leaid = 3620580, and the citywide row has is_aggregate = TRUE. The API adds both fields to its rows. The citywide row's draws are the districts' draws summed draw-wise, so adding it to its districts counts the city twice.

Quick start (DuckDB)

INSTALL httpfs; LOAD httpfs;

-- Summary estimates: download/attach summary.duckdb, then query arrest_summary / state_summary.
-- Raw draws for one slice (TX, Black males, default model, 2023-24), pinned to 2026.1:
SELECT *
FROM read_parquet(
  'hf://datasets/civilytics/crdc-school-arrest-rates@civilytics-crdc-arrests-2026.1/parquet/model_id=unified_m2_mod/YEAR=23-24/LEA_STATE=TX/*.parquet'
)
WHERE RACE = 'BL' AND SEX = 'M'
LIMIT 20;

summary.duckdb direct URL, pinned to 2026.1: https://huggingface.co/datasets/civilytics/crdc-school-arrest-rates/resolve/civilytics-crdc-arrests-2026.1/summary.duckdb

Live API & documentation

• API: https://crdc-api.civilytics.org (OpenAPI/Swagger at /api/v1/__docs__/, machine guide at /api/v1/llms.txt). Every data endpoint takes release=2025.1 or release=2026.1; the default is the latest.
• Docs + data dictionary: https://pages.civilytics.org/crdc-arrests

Source data

U.S. Department of Education, Office for Civil Rights — Civil Rights Data Collection (CRDC), 2015-16, 2017-18, 2021-22 and 2023-24; joined to NCES Common Core of Data (CCD) school and district directories for names and geography, through the CRDC Appendix K crosswalk. Sample restrictions (e.g., enrollment ≥ 30) and the rules for CRDC reserve codes are documented in the data dictionary.

Citation

Knowles, J. E., & Miller, H. (2025). Equity Analysis at a Large Scale: Using Small Area Estimation to Get the Most from the CRDC School Arrest Data. Civilytics Consulting. https://www.civilytics.com/portfolio/equity-analysis-at-a-large-scale/

Name the data release you used alongside it, for example civilytics-crdc-arrests-2026.1.

License

Released under the Open Data Commons Attribution License (ODC-BY 1.0) — you may share and adapt the data provided you attribute the source (the citation above). Underlying CRDC data are U.S. federal government public records.