CRDC Arrest API — Data Dictionary
Releases
Each release is a fixed vintage of the estimates. Numbers within a
release never change, and a later release can revise earlier years.
Every data endpoint takes release= with a release id such
as 2025.1, or latest, which is the default.
Every response names the release that answered it in
meta.data_release. Pin release= when results
need to be reproducible.
GET /api/v1/releases lists each release with its years,
its default model and its change notes. Draws located through
/draws are pinned to the release’s Hugging Face tag,
civilytics-crdc-arrests-{release}.
| release | status | years | m1, m2 cover |
|---|---|---|---|
| 2026.1 | preliminary update | 15-16, 17-18, 21-22, 23-24 | 23-24 |
| 2025.1 | the white paper’s release, vetted for it | 15-16, 17-18, 21-22 | 21-22 |
2025.1 is the white paper’s release, and it was vetted for the paper.
2026.1 is an update, and its estimates are preliminary, because they
have not yet been vetted the same way. Pin release=2025.1
for the paper’s numbers, and see Citation. Release 2026.1 below lists what changed and
the release’s known limits.
Tables
arrest_summary (LEA grain)
| column | meaning |
|---|---|
| LEAID | NCES local education agency id (7-char) |
| lea_name, state_name, lat, lon | from CCD district directory |
| LEA_STATE | USPS state abbreviation |
| YEAR | a school year in the release (see Releases) |
| RACE | AM, BL, HI, WH |
| SEX | F, M |
| model_id | unified_m1..unified_m5 / stratified_m1..stratified_m5 (+ “_mod”) |
| subgroup_id | unified: == model_id; stratified: RACE_SEX (e.g. BL_M) |
| n_draws | posterior draws summarized |
| stu_enroll | binomial denominator: enrollment of the cell’s schools that reported arrests. From 2026.1 a school with a missing report is left out (see Reserve codes); 2025.1 counted every school |
| observed_arrests | raw CRDC arrests in the same schools (reference) |
| count_median/mean/sd | posterior predicted arrest COUNT |
| count_lower_{50,80,95}/upper | HPD interval bounds (counts) |
| rate_median/mean | count / stu_enroll |
| rate_lower_{50,80,95}/upper | count bounds / stu_enroll |
| is_aggregate | API field. TRUE for a row that totals other districts (from 2026.1, New York City’s citywide row); FALSE otherwise, and always FALSE in 2025.1 |
| parent_leaid | API field, present only on a district that an aggregate row totals: the aggregate’s LEAID (3620580 for New York City’s geographic districts) |
state_summary (state grain)
As above, keyed by LEA_STATE (no LEAID/geo). Computed by draw-wise population aggregation: within each draw, sum counts and enrollment across LEAs, compute the rate, then summarize across draws. NOT the (1|LEA_STATE) random effect.
district_dim (one row per LEA)
LEAID, lea_name, LEA_STATE, state_name, lat, lon and enrollment, plus
parent_leaid and is_aggregate as above.
/districts reads it.
Codes
- RACE: AM=American Indian/Alaska Native, BL=Black, HI=Hispanic, WH=White.
- SEX: F, M.
- Models: see /models, which lists the registry of the requested release. Default model when unspecified = unified_m2; the stratified fits (stratified_m1..stratified_m5) must be requested explicitly via model=. m1 and m2 fit the release’s most recent year alone. m3–m5 pool every year in the release: 2025.1 labels that sample “three-year”, and later releases label it “pooled”.
- The default model covers only the most recent year. In 2026.1 a
request for 15-16, 17-18 or 21-22 needs m3, m4 or m5, or
release=2025.1. - m2, m4 and m5 use referral covariates and drop the rows missing one. In 2026.1, m2 drops none; m4 and m5 drop 8 LEA-years in the pooled sample that have no referral rate, and m5 drops one more whose referral total is missing. The 8 have m1 and m3 estimates only; the ninth, Quincy SD 172 (IL) in 2015-16, also has m4.
Unified vs. stratified
Unified unified_m* fit all groups jointly (RACE*SEX); a
row is the unified model’s prediction for that cell. Stratified
stratified_m* fit each RACE×SEX separately; query by
race/sex resolves to that group’s fit (subgroup_id=RACE_SEX).
Intervals
HPD (smallest-width) at the requested mass (50/80/95). Rate bounds are count bounds divided by enrollment.
Sample restrictions (from models.md)
At least 30 students in the district’s arrest-reporting cells; schools whose highest grade is 7 or above; arrests/referrals capped at enrollment; RACE in {AM,BL,HI,WH}; TOTAL not modeled for subgroups; PR excluded.
Reserve codes and missing reports
CRDC marks a cell it has no count for with a negative reserve code. From 2026.1:
-9(not applicable, or skipped) is a structural zero. 2023-24 made the arrest instance items required, so in that wave-9marks most schools with no arrests.- In 2023-24,
-3(a skip-logic failure in the submission system) is 0 where the school reports zero instances for that disability group, and missing otherwise. Earlier waves publish no instance counts, so-3is missing there. - Every other code is missing:
-5(action plan),-6(force certification),-7, and-10to-13. A missing cell leaves both the arrest count andstu_enroll. -10is suppression under Executive Order 14168. It appears only in nonbinary cells, which are not modeled.- An LEA with every cell missing in a year is not modeled for that
year. An LEA with some cells missing is modeled on the cells that
reported; the pooled sample has 64 such LEA-years
(
share_enroll_reportedbelow 1 inmultiyear_data.parquet).
2025.1 read every reserve code as 0, so an unreported school counted as a school with no arrests.
New York City
CRDC files New York City as one LEA, 3620580. The NCES
Common Core of Data lists its schools under 33 geographic-district
LEAIDs, so a direct join loses the city. From 2026.1 the join goes
through the CRDC Appendix K crosswalk, and the models fit the 33
geographic districts.
The city appears in 2021-22 and 2023-24 only. Its 2015-16 schools
report under an action plan (-5) and nearly all its 2017-18
schools under force certification (-6), so those years are
missing.
From 2026.1, New York City appears both as its geographic districts
and as one citywide row (LEAID = 3620580,
is_aggregate = TRUE), whose draws are the districts’ draws
summed draw-wise. Adding the citywide row to its districts counts the
city twice. /districts returns the same two fields.
- The citywide row is never modeled itself.
- A race, sex and model gets a citywide row only when all 33 districts have draws in it. In 2026.1 every one does.
- Its
observed_arrestsandstu_enrollare NA if any district’s are.
Release 2026.1
2026.1 adds 2023-24, and it changes the 2015-16 to 2021-22 numbers for two reasons: unreported schools are now missing rather than zero (see Reserve codes), and New York City is in. The pooled models now fit four waves, so their estimates moved for every year.
Convergence
The usual targets are R-hat below 1.01 and bulk effective sample size (ESS) above 400 for four chains. No fit has a divergent transition or a maximum-tree-depth hit.
unified_m3 was refit with 10,000 iterations and 2,000
warmup, for 4 chains × 8,000 post-warmup draws. Every parameter has bulk
ESS of at least 719 and tail ESS of at least 1,127, and the minimum BFMI
is 0.75. 17,358 of its 17,363 parameters have R-hat below 1.01. The
other five are district intercepts:
| LEAID | district | R-hat | bulk ESS |
|---|---|---|---|
| 1712000 | DeKalb CUSD 428 (IL) | 1.0117 | 936 |
| 1709930 | Chicago Public Schools Dist 299 (IL) | 1.0113 | 980 |
| 5103130 | Prince William County Public Schools (VA) | 1.0106 | 1,209 |
| 5103240 | Richmond City Public Schools (VA) | 1.0105 | 1,178 |
| 1734510 | Rockford SD 205 (IL) | 1.0102 | 1,076 |
The other nine fits ran the iterations below per chain, half of them
warmup. Unified m4 and m5 and stratified m3 to m5 keep every second
post-warmup draw (thin = 2). Each stratified model is eight
fits, one per race and sex.
| model | fits | iterations | max R-hat | parameters with R-hat > 1.01 | min bulk ESS | min tail ESS | min BFMI |
|---|---|---|---|---|---|---|---|
| unified_m1 | 1 | 3,500 | 1.021 | 36 of 16,322 | 341 | 648 | 0.75 |
| unified_m2 | 1 | 3,500 | 1.017 | 21 of 16,323 | 414 | 1,102 | 0.83 |
| unified_m4 | 1 | 4,000 | 1.010 | 1 of 17,364 | 394 | 845 | 1.21 |
| unified_m5 | 1 | 4,000 | 1.019 | 45 of 17,365 | 326 | 714 | 1.12 |
| stratified_m1 | 8 | 2,000 | 1.019 | 27 of 107,284 | 283 | 535 | 0.86 |
| stratified_m2 | 8 | 3,500 | 1.013 | 1 of 107,300 | 500 | 1,021 | 0.86 |
| stratified_m3 | 8 | 3,500 | 1.020 | 178 of 125,199 | 183 | 330 | 1.15 |
| stratified_m4 | 8 | 4,300 | 1.012 | 1 of 125,207 | 596 | 1,138 | 1.18 |
| stratified_m5 | 8 | 4,000 | 1.022 | 30 of 125,215 | 521 | 930 | 1.20 |
Likely reporting errors, kept as reported
Three districts report arrest counts that look like reporting errors. They are in the data as CRDC published them.
| district | 2015-16 | 2017-18 | 2021-22 | 2023-24 |
|---|---|---|---|---|
| Derby (KS), 2005460 | — | 110 | 1,883 | 3 |
| DeKalb CUSD 428 (IL), 1712000 | 0 | 8 | 1,464 | 2 |
| Chicago Public Schools Dist 299 (IL), 1709930 | 1,028 | 25 | 1 | 479 |
Derby’s and DeKalb’s 2021-22 arrests are within two of their referrals (1,881 and 1,463), the pattern of one number entered in both fields. Derby’s 2015-16 arrest data are missing.
The pooled models give each district one intercept across all four
waves, so a misreported year moves that district’s estimate in every
wave. For DeKalb in 2023-24, unified_m3 estimates 311
arrests (95% interval 281 to 345) against 2 reported. The models that
use the referral rate estimate 14 (m4) and 15 (m5), and the recent-year
models, which see 2023-24 alone, estimate 1. The screen that found these
districts flags 134 LEA-years across the four waves: data/qa/outlier_lea_years.csv,
with each unified model’s estimates in outlier_lea_years_estimates.csv.
Citation
Knowles, J. & Miller, H. (2025). Equity Analysis at a Large Scale: Using Small Area Estimation to Get the Most from the CRDC School Arrest Data. Civilytics. https://www.civilytics.com/portfolio/equity-analysis-at-a-large-scale/
Name the data release you used, for example
civilytics-crdc-arrests-2026.1.