ICE arrests joined with detention stays data release

[Deportation Data Updates]

We have released a new dataset that joins ICE arrests with their subsequent detention stays. The dataset allows users to look at arrests together with their detention consequences in a single tooland to download them in a single dataset/spreadsheet. The dataset has one row per arrest — the same coverage as the arrests dataset, from October 1, 2022 through March 10, 2026 — but now also includes information about subsequent detention, including book-in and book-out dates and details on the first, longest, and last stints in each stay in ICE detention.

We have also released two improvements to existing datasets:

1. A new variable in the arrests table, apprehension_state_filled_in, which fills in state where missing, reducing missingness from 15% to 0.5%.
   
2. Improved duplicates treatment across arrests, detention stints, and detention stays datasets. We now include a variable, duplicates_drop_row, to indicate which rows we recommend excluding as likely duplicates.

To build the new arrests-with-stays dataset, we first collapsed arrests that occurred within 24 hours of each other for the same noncitizen into a single arrest episode, keeping the first arrest in each episode (flagged with the duplicate_episode_first variable). We then looked for a detention stay that began in a window around each arrest (five days before to ten days after the book-in, reflecting uncertainty in these dates), and, when more than one arrest fell in that window, we kept the closest. Arrests that did not result in a book-in (and therefore likely resulted in swift release) are kept as well, flagged as false' in thehas_detention` variable. We describe the matching steps in full in the codebook, and we publish the code so you can inspect or adapt it.

One important caveat: many detention stays begin without an ICE arrest—most often at the border, following a CBP arrest—and those stays are not included here. If your focus is on detention rather than on ICE arrests and their consequences, we recommend using the detention stays or detention stints datasets instead.

We welcome suggestions and corrections.