Domain Atlas / Housing & homelessness services
LA County Homelessness Prevention Unit
Work with this case in the PAN Lab ↗
In a Los Angeles County pilot, 335 people who enrolled in the voluntary Homelessness Prevention Unit were reported to be 71% less likely than a regression-adjusted comparison group of 1,285 eligible non-enrollees to enter a homeless shelter or have street-outreach contact within 18 months; the California Policy Lab describes this as an association not yet shown to be causal, pending a randomized controlled trial with results expected in 2027.[3]
What happened
Los Angeles County's Homelessness Prevention Unit (HPU) runs on a predictive risk model built by the California Policy Lab (CPL) at UCLA and operated inside the county Department of Health Services' Housing for Health division, with data linked by the county Chief Information Office. The approach began as a 2019 proof-of-concept, developed with the University of Chicago Poverty Lab, that predicted first-time homelessness among single adults; the county Board of Supervisors approved and funded the program in 2020 with roughly $26 million in federal COVID-relief money, and the deployed unit launched in 2021. The fielded model is reported to use about 580 factors to score a person's risk of becoming homeless within roughly twelve months, drawing on de-identified administrative data — emergency-room visits, behavioral-health care, arrests and probation, benefit enrollment and applications, and prior homeless-services contact. Operational coverage names three departments feeding the deployed model (public social services, mental health, and health services); the seven-agency data linkage described in the record belongs to the 2019 research layer. The model ranks roughly 90,000 prediction-eligible residents from 1 to 90,000 and targets the top ~10,000 for proactive outreach; about 7% of the full 90,000 became homeless within 18 months, versus about 24% of the high-risk group. Participation is entirely voluntary: case managers make three to four contact attempts by phone, letter, and email, and about nine in ten of those successfully reached accept help — enrollment, not scoring, is the documented hard step, rising from 21% to 35% after a dedicated outreach team and standardized discharge were added. Enrollees received an average of $6,469 in flexible financial assistance plus about six months of small-caseload case management. In a pilot early-outcomes analysis (May 2022–February 2023), CPL reported that 335 enrolled participants were 71% less likely to enter a homeless shelter or have street-outreach contact within 18 months than a regression-adjusted comparison group of 1,285 eligible non-enrollees — an association CPL explicitly states is not yet causal, to be resolved by a formal randomized controlled trial with results expected in 2027. CPL's November 2024 equity audit, on a test population of 47,582, examined false-negative rates by race, ethnicity, and gender and found no evidence of systematic exclusion: the model misses a majority of people who later become homeless, but does so roughly consistently across groups (about 56% for Black individuals versus roughly 63–65% for others), and if anything identifies Black individuals slightly better. The program had served 1,498 people to date as of July 2025 (712 households in 2024), with about 86% of completers retaining or entering permanent housing. Its federal COVID-relief funding was scheduled to end in 2026, and observers questioned the value of identifying more people in need without additional housing resources to match.
The sociotechnical reading
Almost every predictive tool in this Atlas points a score at an adverse decision — screen in, investigate, deny, cut hours — so its governance is a fight over who was wrongly selected. HPU is the case that inverts the sign. The score points at an offer, not a sanction; participation is voluntary; the model never denies anyone anything; and its dominant error is not a false positive but a false negative — a high-risk person it misses or that outreach never reaches. The audit that would be a scandal on a punitive tool is almost reassuring here: the miss rate is above half, but it is roughly equitable across race and gender, so on this deployment equity is not the failure — coverage is. That flips where the governance work lives. It is not in contesting who was flagged; it is in defending who was left off the list. The load-bearing structures are the state-feedback channel that keeps outreach looking past the ranked list, the memory loop in which yesterday's records become today's rank, and the slow, periodic check — an equity audit and a trial not due to report causally until 2027 — that is the only thing watching who the model keeps missing. And accuracy is not the lever: a sharper ranker still depends on reaching a person who can say no, so the binding constraints are reach and enrollment, not the score. The distinct lesson the Atlas draws here is that when the dominant error is a missed person rather than a wrongful flag, governance shifts from auditing selection to defending coverage — and that a genuinely benevolent deployment can still be governed less by its model than by its clocks: a funding clock set to run out in 2026 and an evidence clock whose most important number is not yet known.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.