Questions it can answer
- Do shelters send a nightly count, and which ones stop?
- How often does a reported count match the shelter's own log, and in which direction is it wrong?
- Does a bed held for someone on the way get used, released or lost?
- Does any of this differ between arms: a reminder against none, holds against no holds, urban against rural?
How a study runs
- Enroll. A shelter operator picks its projects from HUD's inventory on the enrollment page. The study team issues each shelter a private reporting link.
- Assign. The team names the arms and a seed. Shelters are shuffled with that seed and dealt to the arms in turn, so arm sizes differ by at most one. The seed and each shelter's place in the draw are stored, and the draw can be re-run and checked. A shelter placed in an arm is never moved.
- Run. Staff send the count each evening. Outreach workers with a holding link can hold a bed against a reported count for 90 minutes and receive a code.
- Audit. On nights the team chooses, staff enter what their own log shows at that moment. The platform stores it beside the count that stood as reported.
- Read. The team's console shows response by arm, audit results by arm and hold outcomes, and exports each log as a file.
What is measured
| Measure | Definition |
|---|---|
| Response rate | Nights on which a shelter sent a count, divided by nights in the window, by arm |
| Over-reported | Audits where the reported open count was higher than the log. A person sent there finds no bed. |
| Under-reported | Audits where the reported open count was lower than the log. A free bed went unseen. |
| No count standing | Audits taken when the shelter had sent nothing in the last 18 hours |
| Hold outcomes | Arrived, released by the shelter, cancelled by the referrer, or lapsed after 90 minutes |
The data you get
| File | One row per | Columns |
|---|---|---|
| assignments | Shelter in the study | site_id, label, arm, seed, position, assigned_at |
| reports | Count sent | site_id, arm, project_id, open_beds, offline_beds, reported_at, source |
| audits | Audit check | site_id, arm, project_id, reported_open, reported_at, actual_open, checked_at |
| holds | Hold placed | project_id, beds, created_at, expires_at, closed_at, outcome |
Project ids are the ones in HUD's inventory, so every file joins to the public record: bed type, household type, funders, county and congressional district. The requirements vocabulary is published at /data/ontology.json.
What it does not collect
No name, date of birth, identifier or free text about a guest is collected, and no form has a field for one. A hold is a code and a number of beds. A study that follows people, such as a check-in scan that measures length of stay, runs on the university's systems under its own review board. This platform supplies the shelter side that such a study joins to.