Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

LA County Homelessness Prevention Unit

The help you have to be found for: a homelessness-prevention model

A model ranks tens of thousands of residents by their risk of losing housing and hands outreach workers a short list to go find — an offer of cash and help, never a denial, and about nine in ten of the people actually reached say yes. Modeled on LA County's Homelessness Prevention Unit. The twist is the error that matters runs the other way: the model misses more than half of the people who later become homeless, and the hardest step is not scoring but reaching them. Watch what happens to the people the list left off it.

Stylized model of a documented deploymentHousing & homelessness services

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the HPU-class homelessness-prevention model network: 5 components and 12 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 6 assumed · 1 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

  • assumed

    This models the positive-pole predictive-prevention pattern documented in the LA County Homelessness Prevention Unit case file — not a reconstruction of the actual model or program.

  • baseline

    The harm mode here is omission, not wrongful denial: the model recommends whom to reach and never issues a denial or benefit cutoff, participation is voluntary, and about nine in ten of those reached accept. The dominant error is a false negative — a high-risk person the model misses or that outreach never reaches — so contamination here means the model's blind spots propagating into practice, not a punitive flag.

  • assumed

    The memory loop from de-identified multi-agency records into each score is present at baseline: the score is computed from records the resident never provided for this purpose, and prior service contact raises the risk rank.

  • assumed

    Peer pathways are authored on both signs: outreach practices spread across the team and one model ranks the whole population (so its blind spots are correlated across the cohort), while the inhibiting checks are the defining absence — the light peer second-look runs low, and no independent model re-checks who the list left off, so the majority the model misses goes uncaught between slow, periodic equity audits.

  • assumed

    The independent evaluator (a randomized trial with results expected in 2027, plus a false-negative equity audit) is drawn with two separate oversight inflows — fielded scores from the model, and realized outcomes read from the linked county data — because reading the ground truth independently of the score is what makes the audit an audit. Both inflows are slow and periodic, not per-case second reads, which is why the headline early result is still associational rather than causal.

  • assumed

    The ranked outreach list is drawn as a mediating artifact on the model → outreach pathway, reflecting the case file's account of a system whose output was a prioritized list of people to reach. It carries no flow of its own and does not affect the dynamics; the score routes an offer of help rather than issuing a decision, so the adoption baseline is moderate rather than saturating.

  • assumed

    Differential harm to the people the model ranks, and homelessness itself, are not modeled here; the Lab models institutional propagation only, and those outcomes are documented in the case file and measured outside a diagram like this one. The audited false-negative rate is roughly equitable across race, ethnicity, and gender, with a slightly stronger identification of Black individuals; no differential client-harm figure is asserted here.

What this example does not show

  • Homelessness itself, and differential harm to the people the model ranks, are not modeled here; the Lab models institutional propagation only, and those outcomes are documented in the case file and measured outside any diagram like this one.
  • The reported 71% lower use of shelter or street outreach is associational, drawn from a regression-adjusted comparison of 335 enrollees against 1,285 non-enrollees rather than from the randomized trial whose results are expected in 2027; nothing here should be read as a proven causal effect of the program.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

  • 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.

    empirical
    • Government evaluation Blackwell, Caprara, Rountree, Casey, Vanderford, Battis, Early Outcomes from the Los Angeles County Homelessness Prevention Unit (California Policy Lab, UCLA, 2025) https://capolicylab.org/early-outcomes-from-the-los-angeles-county-homelessness-prevention-unit/
    • Government County of Los Angeles, New Report: Early Signs of Success from LA County's Homelessness Prevention Pilot (2025) https://lacounty.gov/2025/07/10/new-report-early-signs-of-success-from-la-countys-homelessness-prevention-pilot/
    • Reference UCLA Newsroom, Homelessness Prevention Unit participants 71 percent less likely to enter a shelter, California Policy Lab at UCLA finds (2025) https://newsroom.ucla.edu/stories/homeless-prevention-unit-helps-keep-people-off-streets-california-policy-lab-at-ucla

  • The Homelessness Prevention Unit's own November 2024 equity audit, on a test population of 47,582 individuals eligible to be scored, reported false-negative rates ranging from about 56% for Black individuals to roughly 63 to 65% for other groups: the model misses a majority of the people who later become homeless, while performing roughly consistently across race, ethnicity, and gender and identifying Black individuals slightly more strongly.

    empirical
    • Government evaluation California Policy Lab, The Homelessness Prevention Unit: A Proactive Approach to Preventing Homelessness in Los Angeles County (UCLA, 2024) https://capolicylab.org/the-homelessness-prevention-unit-a-proactive-approach-to-preventing-homelessness-in-los-angeles-county/
    • Trade press Fox-Sowell, LA County's New Predictive Model Shows Early Success in Homelessness Prevention Unit (StateScoop, 2025) https://statescoop.com/la-county-ai-predictive-model-reducing-homelessness/

Where this connects

Institutional pressures in this domain

  • Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
  • Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
  • Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
  • Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
  • Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Housing & homelessness services domain page.

Levers available here and the patterns behind them

Documented case histories