Domain Atlas / Housing & homelessness services
Santa Clara County Homelessness Prevention System
In a registered randomized controlled trial of 1,263 imminent-risk applicants (514 treatment, 749 control) run by the University of Notre Dame's evaluation lab, households offered flexible emergency financial assistance averaging about 2,000 dollars, typically one to two months of back rent, through Santa Clara County's homelessness-prevention system were reported 81 percent less likely to become homeless within six months and 73 percent within twelve months; the peer-reviewed article's abstract states the assistance reduced homelessness by 3.8 percentage points from a 4.1 percent base rate, and the researchers conservatively estimated 2.47 dollars in community benefits per net dollar spent.[3]
What happened
Destination: Home, a backbone nonprofit, launched Santa Clara County's Homelessness Prevention System in 2017 with about 1 million dollars in donations, serving roughly 200 households in its first year. Coordinated by Sacred Heart Community Service through a network of community-based partner organizations, the system now runs on roughly 30 million dollars a year, mostly public funds since the County of Santa Clara fully integrated it into its safety net in 2024, and assists about 2,500 households annually (the 30-million-dollar budget, the roughly 6,500-dollar average, and Sacred Heart's coordinating role are reported by CalMatters; Destination: Home's own page states a flat roughly 7,000-dollar average and does not name Sacred Heart or the budget). Since 2017 the program reports assisting more than 31,000 households, representing nearly 44,000 people, with flexible assistance that has covered back rent, deposits, utilities, car repairs, medical bills, and other destabilizing expenses; households can return for repeat assistance, and Destination: Home reports that more than 90 percent remain housed two years later. Those program-scale figures, and the framing of a "first randomized controlled trial of its kind," are Destination: Home's own reporting, not independently audited.
The targeting is not a headline machine-learning system, and keeping that straight is the point of this entry. Eligibility and prioritisation run on an intake vulnerability questionnaire that scores homelessness risk from self-reported factors such as domestic-violence history, prior homelessness, and disability. During the evaluation window, applicants at imminent risk who scored in a middle band (8 to 13 on the vulnerability assessment) and were ineligible for other prevention programs were allocated by lottery rather than first-come-first-served, because demand exceeded available funds. This is a transparent points-based screen, distinct from the supervised machine-learning ranker used by the Los Angeles County Homelessness Prevention Unit (whose causal trial is still pending) and from the separate 2016 Silicon Valley Triage Tool, a predictive model for prioritising people who are already homeless.
The evidence that puts this program in the Atlas is a registered randomized controlled trial run by the University of Notre Dame's Wilson Sheehan Lab for Economic Opportunities (LEO), led by David C. Phillips and James X. Sullivan (AEA RCT Registry AEARCTR-0008261). The trial enrolled 1,263 people (514 treatment, 749 control) from July 2019 to December 2020, randomized individually via SurveyCTO, with homelessness measured through the county Homeless Management Information System; the control group received non-financial services only, while the treatment group added an average of about 2,000 dollars in assistance, typically one to two months of back rent paid directly to landlords, alongside non-financial help such as credit counselling and landlord dispute resolution. The reported effect is large, and its figures come from different documents and should not be fused: Notre Dame and LEO releases state that assisted households were 81 percent less likely to become homeless within six months and 73 percent within twelve; CalMatters reports the underlying rates as 0.9 percent of assisted versus 4.1 percent of unassisted households, which it pairs with a 78 percent relative reduction; and the peer-reviewed article's abstract states that the assistance reduced homelessness by 3.8 percentage points from a 4.1 percent base rate, with effects larger for people with a history of homelessness and no children. The authors conservatively estimate 2.47 dollars in community benefits per net dollar spent. The article of record is Phillips and Sullivan, Review of Economics and Statistics 107(5): 1187 to 1196, published in September 2025, though the findings were first publicized in mid-2023 (one publisher explainer even refers to a "May 2023 issue"); the paper's own abstract claims only "the first evidence from a randomized controlled trial isolating the impact of financial assistance to prevent homelessness," a narrower claim than the program's "first RCT" framing. In 2024 the county integrated the model into its public safety-net systems, and Results for America released a replication toolkit.
The program's sharpest limitation is named by its own evaluators, not by outside critics. Because becoming homeless is statistically rare even among at-risk applicants - about 96 percent of the trial's control group never became homeless without the assistance - accurate risk identification, not the strength of the help, is the binding constraint. Co-author James Sullivan cautions on the record that "precious resources" could go to people who would have stayed housed anyway, diverting funds from shelter and permanent housing; and the UCLA California Policy Lab's Janey Rountree has made the related point (paraphrased in CalMatters) that homelessness is statistically extremely rare and most people at risk stay housed through family and friends, which is what makes accurate targeting so hard. The trial also carries real external-validity caveats: it studied a middle band of risk scores, so effects at other risk levels are unmeasured, and its July 2019 to December 2020 enrollment overlapped COVID-era eviction moratoria and emergency rental assistance, which may have depressed base rates.
The model is now the seed of a live, multi-site experiment. On February 24, 2026, Destination: Home launched Right at Home, a five-year (2026 to 2031) national initiative backed by 77 million dollars from The Audacious Project at TED, Cisco, Sobrato Philanthropies, and the Valhalla Foundation (CalMatters reports nearly 80 million dollars raised), to replicate the model in ten communities. Eight are named - Alaska, the Asheville-Buncombe region of North Carolina, Atlanta, Austin-Travis County, San Mateo County, the Denver-Adams County area, Miami-Dade County, and the Minnesota Tribal Collaborative Pathways to Housing - with two locations pending. Each site is to receive a minimum of 5 million dollars over three years and to begin implementation by January 2027, with a goal of keeping more than 10,000 households housed, and with LEO serving as the common national evidence partner rigorously evaluating every site. Because those jurisdictions are structurally heterogeneous - urban counties, a state coalition, and tribal communities - Right at Home is effectively a test of whether a lever whose strength was measured once, on a middle band, in one county, during a pandemic eviction moratorium, survives the change of setting. The model is also driving a California policy push: San Mateo and Marin counties are launching pilots, San Francisco and Oakland run similar programs (CalMatters reports San Francisco participants under 5 percent homeless within a year versus 8 percent of comparable non-participants, March 2023 to February 2025), and AB 1924, carried by Assemblymember Jesse Gabriel and co-sponsored by the All Home coalition, would require a state homelessness-prevention strategy by July 2027, though with no funding attached.
The sociotechnical reading
The Atlas already carries a careful homelessness-prevention counterexample - the Los Angeles County Homelessness Prevention Unit - and it is worth saying exactly how this one differs, because the difference is the lesson. LA County uses a supervised machine-learning model to rank tens of thousands of residents and surface a short outreach list; its causal trial is still pending, and its dominant failure mode is the miss - the high-risk person the model never ranks or outreach never reaches, a harm by omission. Santa Clara is almost the mirror image on every axis that matters. Its targeting is a transparent points questionnaire, not an opaque model, so nobody is arguing about whether to trust a black box. Its trial is done and published. And its characteristic failure mode is not the miss but the deadweight: spending scarce assistance on households that would have kept their housing anyway, a harm by misallocation rather than omission.
That inversion is what makes this the clearest case in the Atlas of a system where the model is the least interesting part - and where getting that recognition right is the whole point. Read the system as a map and the containment is real and in the right places: the intervention delivers a benefit and never a denial, so there is no wrongful-adverse-action edge to worry about; human discretion is high by design, with caseworkers determining assistance case by case and non-financial help alongside the money, so the correction edge is strong rather than eroding; and a closed measurement loop runs through the county HMIS, which is precisely what made a randomized trial possible in the first place. The result is genuinely good, at 2.47 dollars of community benefit per net dollar. So the governance question is not "trust the model less." It is the one thing all of that leaves untouched: whether the screen is reaching the right people.
Here the system map earns its keep, because it separates two measurements that celebration tends to merge. The effect size answers "does the lever work?" - and it does. It does not answer, and cannot answer, "does the screen select the households that would otherwise have become homeless?" On a rare-event base rate, those are different questions, and the second dominates: when about 96 percent of even an at-risk, middle-band population stays housed without help, most of the money necessarily flows to people who never needed it, and the program's own co-evaluator says so on the record. The trap is that this deadweight is invisible. The people who would have stayed housed do stay housed, so mis-targeting looks exactly like success - the same HMIS loop that shows the lever works will keep returning reassuring outcomes whether or not the target is right. A measured lever, in other words, can quietly stand in for a measured target, and on this base rate that substitution is the failure mode.
Which is why the productive corner is not a sharper questionnaire but a reconciled target. Three moves matter, and none of them touches the model's accuracy. First, stand up the check the effect size skips: reconcile the screen's targeting against measured outcomes, so "the help works" is joined by "the help is reaching the households the base rate says would otherwise fall," a check that on this shape is the binding one and currently does not exist as a standing edge. Second, keep the effect legible for what it was - measured once, on a middle band, in one county, during an eviction moratorium - so a number does not travel to a new place as a property it inherits rather than a result with a setting attached. Third, treat the replication as what it is: a live transfer of one lever design across ten structurally different topologies at once, with the discipline of authorizing each site with its own evaluation attached rather than on the strength of Santa Clara alone - which, to the program's credit, is close to how Right at Home is actually funded, with per-site grants and a common evidence partner evaluating every location. The distinct lesson the Atlas draws here: a measured lever is not a measured target, and on a rare-event base rate the second question dominates the first. A program can be genuinely, measurably effective at the thing it measured and still be spending scarce resources on people who never needed them - and the only way to tell the difference is to check the target as rigorously as the effect, at every site, every time the lever moves to a topology no one has tested it in yet. The honest boundary throughout: served people, and whether any household does or does not lose its home, are not modeled in the Lab, which reads institutional propagation only; every program-scale figure here is Destination: Home's own reporting, and the trial's effect figures are separately sourced and carry their own external-validity caveats.
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.